山东大学：《发育生物学 Developmental Biology》课程教学资源（参考资料）Circular RNAs are a large class of animal RNAs with regulatory potency

ARTICLE doi:10.1038/ nature11928 Circular RNas are a large class of animal RNAs with regulatory potency Sebastian Memczak *, Marvin Jens *, Antigoni Elefsinioti*, Francesca Torti*, Janna Krueger?, Agnieszka Rybak, Luisa Maier, Mackowiak, Lea H Greger Circular RNAs(circRNAs) in animals are an enigmatic class of RNa with unknown function. To explore circRNAs systematically, we sequenced and computationally analysed human, mouse and nematode RNA. We detected thousands of well-expressed, stable circRNAs, often showing tissue/developmental-stage-specific expression Sequence analysis indicated important regulatory functions for circRNAs. We found that a human circRNA, antisense to the cerebellar degeneration-related protein 1 transcript(CDRlas), is densely bound by microRNA (miRNA)effector complexes and harbours 63 conserved binding sites for the ancient miRNA miR-7. Further analyses indicated that CDRlas functions to bind miR-7 in neuronal tissues. Human CDRlas expression in zebrafish impaired midbrain development, similar to knocking down miR-7, suggesting that CDRlas is a miRNA antagonist with a miRNA-binding capacity ten times higher than any other known transcript. Together, our data provide evidence that circRNAs form a large class of post-transcriptional regulators. Numerous circRNAs form by head-to-tail splicing of exons, suggesting previously unrecognized regulatory potential of coding sequences Mature messenger RNAS are linear molecules with 5'and 3 termini computational pipeline can find circRNAs in any genomic region that reflect start and stop of the RNa polymerase on the DNA tem- takes advantage of long(100 nucleotides)reads, and predicts the plate In cells, different RNA molecules are sometimes joined together acceptor and donor splice sites used to link the ends of the RNAs by splicing reactions(trans-splicing), but covalent linkage of the ends We do not rely on paired-end sequencing data or known splice sites of a single RNA molecule to form a circular RNA(circRNA)is usually Using published 2526 and our own sequencing data, our method considered a rareevent circRNAs were discovered in plants and shown reported thousands of circRNAs in human and mouse tissues as well to encode subviral agents!. In unicellular organisms, cirCRNAs mostly as in different developmental stages of Caenorhabditis elegans stem from self-splicing introns of pre-ribosomal RNA, but can also Numerous circRNAs appear to be specifically expressed across tissues rise from protein-coding genes in archaea. In the few unambiguously or developmental stages. We validated these data and showed that validated circRNAs in animals, the spliceosome seems to link the 5 most tested circRNAs are well expressed, stable and circularized and downstream 3 ends of exons within the same transcript'-lo. using the predicted splice sites. circRNA sequences were significantly Perhaps the best known circRNA is antisense to the mRNA transcribed enriched in conserved nucleotides, indicating that circRNAs compete from the SRY (sex-determining region Y) locus and is highly expressed with other RNAs for binding by RNa binding proteins(RBPs)or in testes. Evidence from computational analyses of expression data in miRNAs. We combined biochemical, functional and computational Archaea and Mammalia suggests that circRNAs are more prevalent analyses to show that indeed a known human circRNA, CDRI anti- than previously thought. however, it is unknown whether animal sense(CDRlas), can function as a negative regulator of miR-7,a miRNA with perfect sequence conservation from annelids to human. In comparison to circRNAS, miRNAs are extremely well studied. Together, our data provide evidence that circRNAs form an important miRNAs are -21-nucleotide-long non-coding RNAs that guide the class of post-transcriptional regulators ffector protein Argonaute(AGO) to mRNAs of coding genes to press their protein production"-14. In humans, miRNAs directly circRNAs have complex expression patterns regulate expression of most mRNAss-I8 in a diverse range of bio- To comprehensively identify stably expressed circRNAs in animals we ical functions. However, surprisingly little is known about how screened RNA sequencing reads for splice junctions formed by an and if mRNAs can escape regulation by a miRNA. A recently discov- acceptor splice site at the 5 end of an exon and a donor site at a red mechanism for miRNA removal in a sequence-specific manner is downstream 3'end(head-to-tail)(Fig. la). As standard RNA expres- based on target sites acting as decoys or miRNA sponges. RNA with sion profiling enriches for polyadenylated RNAs, we used data gene miRNA binding sites should, if expressed highly enough, sequester rated after ribosomal RNA depletion (ribominus)and random way the miRNA from its target sites. However, all reported mam- priming, Such data were used before to detect scrambled exons in malian miRNA sponges have only one or two binding sites for the same mammals"(see Methods for comparison). However, this approach miRNA and are not highly expressed, limiting their potency was not specifically designed to detect circRNAs and (1)only used To identify circRNAs across animal cells systematically, we screened existing exon-intron annotations, thus missing RNAs transcribed RNA-seq data for circRNAs. Compared to previous approaches our from introns or unannotated transcripts; (2)did not explicitly identi sYstems Biology of Gene Regulatory Elements, Max-Delbrulck-Center for Molecular Medicine, Robert-Rbssle-Strasse 10,13125 Berlin 2Angiogenesis and Cardiovascular Pathology, Max elbrilck-Center for Molecular Medicine, Robert-Rossle-Strasse 10, 13125 Berlin, Germany. RNA Biology and Post-Transcriptional Regulation, Max-Delbriick-Center for Molecular Medicine, Robert- Rdssle-Strasse 10, 13125 Berlin, Germany. signaling Dynamics in Single Cells, Max-Delbriick-Center for Molecular Medicine, Robert-Rbssle-Strasse 10, 13125 Berlin, Germany These authors contributed equally to this work 0 MONTH 2013 VOL 000I NATURE I @2013 Macmillan Publishers Limited. All rights reserved

ARTICLE doi:10.1038/nature11928 Circular RNAs are a large class of animal RNAs with regulatory potency Sebastian Memczak1 *, Marvin Jens1 *, Antigoni Elefsinioti1 *, Francesca Torti1 *, Janna Krueger2 , Agnieszka Rybak1 , Luisa Maier1 , Sebastian D. Mackowiak1 , Lea H. Gregersen3 , Mathias Munschauer3 , Alexander Loewer4 , Ulrike Ziebold1 , Markus Landthaler3 , Christine Kocks1 , Ferdinand le Noble2 & Nikolaus Rajewsky1 Circular RNAs (circRNAs) in animals are an enigmatic class of RNA with unknown function. To explore circRNAs systematically, we sequenced and computationally analysed human, mouse and nematode RNA. We detected thousands of well-expressed, stable circRNAs, often showing tissue/developmental-stage-specific expression. Sequence analysis indicated important regulatory functions for circRNAs. We found that a human circRNA, antisense to the cerebellar degeneration-related protein 1 transcript (CDR1as), is densely bound by microRNA (miRNA) effector complexes and harbours 63 conserved binding sites for the ancient miRNA miR-7. Further analyses indicated that CDR1as functions to bind miR-7 in neuronal tissues. Human CDR1as expression in zebrafish impaired midbrain development, similar to knocking down miR-7, suggesting that CDR1as is a miRNA antagonist with a miRNA-binding capacity ten times higher than any other known transcript. Together, our data provide evidence that circRNAs form a large class of post-transcriptional regulators. Numerous circRNAs form by head-to-tail splicing of exons, suggesting previously unrecognized regulatory potential of coding sequences. Mature messenger RNAs are linear molecules with 59 and 39 termini that reflect start and stop of the RNA polymerase on the DNA tem￾plate. In cells, different RNA molecules are sometimes joined together by splicing reactions (trans-splicing), but covalent linkage of the ends of a single RNA molecule to form a circular RNA (circRNA) is usually considered a rare event. circRNAs were discovered in plants and shown to encode subviral agents1 . In unicellular organisms, circRNAs mostly stem from self-splicing introns of pre-ribosomal RNA2 , but can also arise from protein-coding genes in archaea3 . In the few unambiguously validated circRNAs in animals, the spliceosome seems to link the 59 and downstream 39 ends of exons within the same transcript4–10. Perhaps the best known circRNA is antisense to the mRNA transcribed from the SRY (sex-determining region Y) locus and is highly expressed in testes6 . Evidence from computational analyses of expression data in Archaea and Mammalia suggests that circRNAs are more prevalent than previously thought3,10; however, it is unknown whether animal circRNAs have any biological function. In comparison to circRNAs, miRNAs are extremely well studied. miRNAs are ,21-nucleotide-long non-coding RNAs that guide the effector protein Argonaute (AGO) to mRNAs of coding genes to repress their protein production11–14. In humans, miRNAs directly regulate expression of most mRNAs15–18 in a diverse range of bio￾logical functions. However, surprisingly little is known about how and if mRNAs can escape regulation by a miRNA. A recently discov￾ered mechanism for miRNA removal in a sequence-specific manner is based on target sites acting as decoys or miRNA sponges19,20. RNA with miRNA binding sites should, if expressed highly enough, sequester away the miRNA from its target sites. However, all reported mam￾malian miRNA sponges have only one or two binding sites for the same miRNA and are not highly expressed, limiting their potency21–24. To identify circRNAs across animal cells systematically, we screened RNA-seq data for circRNAs. Compared to previous approaches10 our computational pipeline can find circRNAs in any genomic region, takes advantage of long (,100 nucleotides) reads, and predicts the acceptor and donor splice sites used to link the ends of the RNAs. We do not rely on paired-end sequencing data or known splice sites. Using published10,25,26 and our own sequencing data, our method reported thousands of circRNAs in human and mouse tissues as well as in different developmental stages of Caenorhabditis elegans. Numerous circRNAs appear to be specifically expressed across tissues or developmental stages. We validated these data and showed that most tested circRNAs are well expressed, stable and circularized using the predicted splice sites. circRNA sequences were significantly enriched in conserved nucleotides, indicating that circRNAs compete with other RNAs for binding by RNA binding proteins (RBPs) or miRNAs. We combined biochemical, functional and computational analyses to show that indeed a known human circRNA, CDR1 anti￾sense (CDR1as)9 , can function as a negative regulator of miR-7, a miRNA with perfect sequence conservation from annelids to human. Together, our data provide evidence that circRNAs form an important class of post-transcriptional regulators. circRNAs have complex expression patterns To comprehensively identify stably expressed circRNAs in animals we screened RNA sequencing reads for splice junctions formed by an acceptor splice site at the 59 end of an exon and a donor site at a downstream 39 end (head-to-tail) (Fig. 1a). As standard RNA expres￾sion profiling enriches for polyadenylated RNAs, we used data gene￾rated after ribosomal RNA depletion (ribominus) and random priming. Such data were used before to detect scrambled exons in mammals10 (see Methods for comparison). However, this approach was not specifically designed to detect circRNAs and (1) only used existing exon–intron annotations, thus missing RNAs transcribed from introns or unannotated transcripts; (2) did not explicitly identify *These authors contributed equally to this work. 1 Systems Biology of Gene Regulatory Elements, Max-Delbru¨ ck-Center for Molecular Medicine, Robert-Ro¨ssle-Strasse 10, 13125 Berlin, Germany. 2 Angiogenesis and Cardiovascular Pathology, Max￾Delbru¨ ck-Center for Molecular Medicine, Robert-Ro¨ssle-Strasse 10, 13125 Berlin, Germany. 3 RNA Biology and Post-Transcriptional Regulation, Max-Delbru¨ ck-Center for Molecular Medicine, Robert￾Ro¨ssle-Strasse 10, 13125 Berlin, Germany. 4 Signaling Dynamics in Single Cells, Max-Delbru¨ ck-Center for Molecular Medicine, Robert-Ro¨ssle-Strasse 10, 13125 Berlin, Germany. 00 MONTH 2013 | VOL 000 | NATURE | 1 ©2013 Macmillan Publishers Limited. All rights reserved

RESEARCH ARTICLE alignment and splice-site detection from at least two independent junction-spanning reads(Fig 1b).The expression of genes predicted to give rise to circRNAs was only slightl shifted towards higher expression values( Supplementary Fig. 1d), Linear splicing indicating that circRNAs are not just rare mistakes of the splice- Donor some. We also identified 1, 903 circRNAs in mouse(brains, fetal head, differentiation-induced embryonic stem cells; Supplementary Fig. le)281 of these mapped to human circRNAs(Supplem tary Fig. 1f). To explore whether circRNAs exist in other animal clades, we used sequencing data that we produced from various C. elegan Oocyte 1-cell embryo d Human circRNAs developmental stages(Stoeckius, M et al, manuscript in preparation) Methods)and detected 724 circRNAs, with at least two independent reads(Fig. 1c). Numerous circRNAs seem to be specifically expressed in a cell type or developmental stage(Fig. Ib, c and Supplementary Fig. le). For example, hsa-circRNA 2149 is supported by 13 unique, head-to-tail spanning reads in CD19 leukocytes but is not detected in CD34 leukocytes(which were sequenced at comparable depth; Supplemen K27 tary Table 1), neutrophils or HEK293 cells. Analogously, a number of C elegans(/24) (1.602 inside coding transcripts) nematode circRNAs seem to be expressed in oocytes but absent in f hird codon position I-or 2-cell embryos CDR1 gene catalogue of non-coding RNAs27-29 oing the RefSeq database and a 5% of human circRNAs align Long ncRNA PVT1 sense to known genes. Their splice sites typically span one to five exons(s ntary Fig. 1g) and overlap coding exons(84%),but ZRANBT only in 65% of these cases are both splice sites that participate in the circularization known splice sites(Supplementary Table 2), demon- Consere of\? cos strating the advantage of our strategy. 10% of all circRNAs align antisense to known transcripts, smaller fractions align to UTRS, introns, unannotated regions of the genome(Fig. 1d). Examples of Figure 1 Detection, classification and evolutionary conservation of human circRNAs are shown in Fig. le quentially to the genome for linear( top) but in reversed orientation for head. sea We analysed sequence conservation within circRNAS. As genomic circRNAs. a, The termini of junction-spanning reads(anchors)align quence is subject to different degrees of evolutionary selection, to-tail spliced reads(bottom). Spliced reads must distribute completely to depending on function, we studied three subtypes of circRNAs anchors, flanked by AG/GU (Methods). b, c, circRNAs in human cell types Intergenic and a few intronic circRNAs display a mild but significant (b)and nematode stages(c).d, Genomic origin of human circRNAs. A total of enrichment of conserved nucleotides(Supplementary Fig. Ih, i). 6% of circRNAs overlap known transcripts. e, Examples of human circRNAs. To analyse circRNAs composed of coding sequence and thus high The AFFI intron is spliced out( Supplementary Fig 2e) Sequence conservation: overall conservation, we selected 223 human circRNAs with circular lacental mammals phyloP (Methods), scale bar, 200 nucleotides f a total of 223 human coding sequence circRNAs with mouse orthologues orthologues in mouse(Methods)and entirely composed of coding (green)and controls(black) with matched conservation level(inset: mean sequence Control(linear)exons were randomly selected to match the onservation for each codon position(grey), controls(black); x axis, codon level of conservation observed in first and second codon positions positions y axis, placental mammals phyloP score; see also Methods and (Methods, Fig. If inset and Supplementary Fig. Ik for conservation served(P<4r,i,k). Third codon positions are significantly more of the remaining coding sequence(CDS). circRNAs with conserved ircularization were significantly more conserved in the third codon position than controls, indicating evolutionary constraints at the nuc- the splice sites used for circularization; and (3)assumed that each pair leotide level, in addition to selection at the protein level(Fig. If and of mates in paired-end sequencing derives from the same RNA mole- Supplementary Fig. 1j, k). In summary, we have confidently identified ule. To search in a more unbiased way for circRNAs, we designed a large number of circRNAs with complex expression patterns, which algorithm(Methods) that identifies linear and circular splicing derive often but not always from coding exons Sequence conservation events in ribominus data. First, we filtered out reads that aligned con- suggests that at least a subset contains functional sequence elements tiguously to the genome, retaining the spliced reads. Next, we mapped the terminal parts of each candidate read independently to the genome Characterization of 50 predicted circRNAS to find ue an tions. Finally, we demanded that(1)anchor We experimentally tested our circRNA predictions in HEK293 cells alignments can be extended such that the original read sequence Head-to-tail splicing was assayed by quantitative polymerase chain aligns completely, and( 2)the inferred breakpoint is flanked by GU/ reaction( qPCR) after reverse transcription, with divergent primers AG splice signals. Non-unique mappings and ambiguous breakpoints and Sanger sequencing( Fig. 2a, b). Predicted head-to-tail junctions were discarded. We detected circularization splicing from the reversed of 19 out of 23 randomly chosen circRNAs(83%)could be validated, (head-to-tail) orientation of the anchor alignments(Fig. la). Our demonstrating high accuracy of our predictions(Table 1). In contrast, method also recovered tens of thousands of known linear splicing 5 out of7(71%)candidates exclusively predicted in leukocytes could not events(Methods and Supplementary Fig. la, b). We estimated sen- be detected in HEK293 cells, validating cell-type-specific expression. sitivity(75%)and false-discoveof real sequencing data (ang simulated Head-to-tail splicing could be produced by trans-splicing or geno- eads and various permutations of Methods and mic rearrangements. To rule out these possibilities as well as potentia upplementary Fig. Ic). However, the efficiency of ribominus pro- PCR artefacts, we successfully validated the insensitivity of human tocols to extract and sequence circRNAs is limited, reducing overall circRNA candidates to digestion with RNase R-an exonuclease that degrades linear RNA molecules -by northern blotting with probe We generated ribominus data for HEK293 cells and, combined which span the head-to-tail junctions(Fig. 2c). We quantified RNase with human leukocyte data, detected 1, 950 circRNAs with support Rresistance for 21 candidates with confirmed head-to-tail splicing by 2I NATURE I VOL 00000 MONTH 2013 @2013 Macmillan Publishers Limited. All rights reserved

the splice sites used for circularization; and (3) assumed that each pair of mates in paired-end sequencing derives from the same RNA mole￾cule. To search in a more unbiased way for circRNAs, we designed an algorithm (Methods) that identifies linear and circular splicing events in ribominus data. First, we filtered out reads that aligned con￾tiguously to the genome, retaining the spliced reads. Next, we mapped the terminal parts of each candidate read independently to the genome to find unique anchor positions. Finally, we demanded that (1) anchor alignments can be extended such that the original read sequence aligns completely, and (2) the inferred breakpoint is flanked by GU/ AG splice signals. Non-unique mappings and ambiguous breakpoints were discarded. We detected circularization splicing from the reversed (head-to-tail) orientation of the anchor alignments (Fig. 1a). Our method also recovered tens of thousands of known linear splicing events (Methods and Supplementary Fig. 1a, b). We estimated sen￾sitivity (.75%) and false-discovery rate (FDR ,0.2%) using simulated reads and various permutations of real sequencing data (Methods and Supplementary Fig. 1c). However, the efficiency of ribominus pro￾tocols to extract and sequence circRNAs is limited, reducing overall sensitivity. We generated ribominus data for HEK293 cells and, combined with human leukocyte data10, detected 1,950 circRNAs with support from at least two independent junction-spanning reads (Fig. 1b). The expression of genes predicted to give rise to circRNAs was only slightly shifted towards higher expression values (Supplementary Fig. 1d), indicating that circRNAs are not just rare mistakes of the spliceo￾some. We also identified 1,903 circRNAs in mouse (brains, fetal head, differentiation-induced embryonic stem cells; Supplementary Fig. 1e)25,26; 81 of these mapped to human circRNAs (Supplemen￾tary Fig. 1f). To explore whether circRNAs exist in other animal clades, we used sequencing data that we produced from various C. elegans developmental stages (Stoeckius, M.et al., manuscript in preparation) (Methods) and detected 724 circRNAs, with at least two independent reads (Fig. 1c). Numerous circRNAs seem to be specifically expressed in a cell type or developmental stage (Fig. 1b, c and Supplementary Fig. 1e). For example, hsa-circRNA 2149 is supported by 13 unique, head-to-tail spanning reads in CD191 leukocytes but is not detected in CD341 leukocytes (which were sequenced at comparable depth; Supplemen￾tary Table 1), neutrophils or HEK293 cells. Analogously, a number of nematode circRNAs seem to be expressed in oocytes but absent in 1- or 2-cell embryos. We annotated human circRNAs using the RefSeq database and a catalogue of non-coding RNAs27–29. 85% of human circRNAs align sense to known genes. Their splice sites typically span one to five exons (Supplementary Fig. 1g) and overlap coding exons (84%), but only in 65% of these cases are both splice sites that participate in the circularization known splice sites (Supplementary Table 2), demon￾strating the advantage of our strategy. 10% of all circRNAs align antisense to known transcripts, smaller fractions align to UTRs, introns, unannotated regions of the genome (Fig. 1d). Examples of human circRNAs are shown in Fig. 1e. We analysed sequence conservation within circRNAs. As genomic sequence is subject to different degrees of evolutionary selection, depending on function, we studied three subtypes of circRNAs. Intergenic and a few intronic circRNAs display a mild but significant enrichment of conserved nucleotides (Supplementary Fig. 1h, i). To analyse circRNAs composed of coding sequence and thus high overall conservation, we selected 223 human circRNAs with circular orthologues in mouse (Methods) and entirely composed of coding sequence. Control (linear) exons were randomly selected to match the level of conservation observed in first and second codon positions (Methods, Fig. 1f inset and Supplementary Fig. 1k for conservation of the remaining coding sequence (CDS)). circRNAs with conserved circularization were significantly more conserved in the third codon position than controls, indicating evolutionary constraints at the nuc￾leotide level, in addition to selection at the protein level (Fig. 1f and Supplementary Fig. 1j, k). In summary, we have confidently identified a large number of circRNAs with complex expression patterns, which derive often but not always from coding exons. Sequence conservation suggests that at least a subset contains functional sequence elements. Characterization of 50 predicted circRNAs We experimentally tested our circRNA predictions in HEK293 cells. Head-to-tail splicing was assayed by quantitative polymerase chain reaction (qPCR) after reverse transcription, with divergent primers and Sanger sequencing (Fig. 2a, b). Predicted head-to-tail junctions of 19 out of 23 randomly chosen circRNAs (83%) could be validated, demonstrating high accuracy of our predictions (Table 1). In contrast, 5 out of 7 (71%) candidates exclusively predicted in leukocytes could not be detected in HEK293 cells, validating cell-type-specific expression. Head-to-tail splicing could be produced by trans-splicing or geno￾mic rearrangements. To rule out these possibilities as well as potential PCR artefacts, we successfully validated the insensitivity of human circRNA candidates to digestion with RNase R—an exonuclease that degrades linear RNA molecules30—by northern blotting with probes which span the head-to-tail junctions (Fig. 2c). We quantified RNase R resistance for 21 candidates with confirmed head-to-tail splicing by AG GT Circularization 5′ anchor Acceptor Donor AG Acceptor GT Donor Spliced read a circRNA Linear splicing Anchor alignment and splice-site detection Sperm 2-cell embryo Oocyte 1-cell embryo b c Human (1,950) CD19+ HEK293 CD34+ Neutrophils C. elegans (724) (1,602 inside coding transcripts) Intergenic ncRNA Antisense Human circRNAs 152 183 90 59 51 30 31 22 20 20 16 16 16 13 5 106 28 79 12 22 89 20 25 3 19 21 194 939 60 333 81 CDS exons 1,000 5′UTR 195 21 3′UTR 79 27 Intronic 80 147 63 204 Other 53 d hsa-circRNA 6 hsa-circRNA 2 CDR1as hsa-circRNA 9 hsa-circRNA 1862 ZRANB1 exon1 Long ncRNA PVT1 exon3 CDR1 gene AFF1 exon4,5 GT GT GT GT GT AG AG AG AG AG Control cons. CDS circRNA Third codon position 123 2 1 0 Conservation *** 0 1 Conservation score Cumulative frequency 0 0.5 1 e f 3′ anchor Conservation 0.5 Intergenic chr4:42,212,391-42,214,180 Figure 1 | Detection, classification and evolutionary conservation of circRNAs. a, The termini of junction-spanning reads (anchors) align sequentially to the genome for linear (top) but in reversed orientation for head￾to-tail spliced reads (bottom). Spliced reads must distribute completely to anchors, flanked by AG/GU (Methods). b, c, circRNAs in human cell types (b) and nematode stages (c). d, Genomic origin of human circRNAs. A total of 96% of circRNAs overlap known transcripts. e, Examples of human circRNAs. TheAFF1 intron is spliced out (Supplementary Fig. 2e). Sequence conservation: placental mammals phyloP score (Methods), scale bar, 200 nucleotides. f, A total of 223 human coding sequence circRNAs with mouse orthologues (green) and controls (black) with matched conservation level (inset: mean conservation for each codon position (grey), controls (black); x axis, codon positions; y axis, placental mammals phyloP score; see also Methods and Supplementary Fig. 1j, k). Third codon positions are significantly more conserved (P , 4 3 10210, Mann–Whitney U-test, n 5 223). RESEARCH ARTICLE 2 | NATURE | VOL 000 | 00 MONTH 2013 ©2013 Macmillan Publishers Limited. All rights reserved

ARTICLE RESEARCH onvergent(4← ficantly enriched compared to coding sequences(P<2.96X 10 3)or3 AGCTTCCT GTGTGGGT Mann-Whitney U-test, n=3, 182)( Supplementary Fig 3a, b) As an extreme case. we discovered that the known human circrNA CCTG CDRlas(ref 9)harboured dozens of conserved miR-7 seed matches. Sanger se To test whether CDRlas is bound by miRNAs, we analysed bio- hemical, transcriptome-wide binding-site data for the miRNA (photoactivatable-ribonucleoside-enhanced crosslinking and immu- 100小;444谷 precipitation) experiments for human AGO(Methods) and ana- hsa-circRNA2 hsa-circRNA3 hsa-circ RNA16 GAPDH lysed them together with published, lower-depth data 2. PAR-CLIP3 RNaseR is based on ultraviolet crosslinking of rNa to protein and subseque equencing of RNA bound to a RBP of interest. The -1. 5-kilobase(kb CDRlas locus stood out in density and number of AGO PAR-CLIP ao.oLLVIgGRNAS RBPs gave virtually no signal. Of note, there is no PAR-CLIP read reads(Fig. 3a), whereas nine combined PAR-CLIP libraries for other 24 h Acto mapping to the sense coding transcript of the CDRI gene, which was originally identified as a target of autoantibodies from patients with paraneoplastic cerebellar degeneration Agarose gel Sequence analysis across 32 vertebrate species revealed that miR-7 hsa-circ RNAs 2 9 16 is the only animal miRNA with conserved seed matches that can hsa-cincRNAs explain the AGo binding along the CDRlas transcript( Methods) Human CDRlas harbours 74 miR-7 seed matches of which 63 are Figure 2 stable transcripts with robust expressic man(hsa) ZRANBI circRNA exemplifies the validation strategy Convergent(divergent) primers detect total(circular)RNAs. Sanger uencing confirms head-to-tail splicing. b, Divergent primers amplify circRNAs in cDNA but not genomic DNA (gDNA). GAPDH, linear control. ze marker in base pairs. c, Northern blots of mock(-)and RNaseR(+) eated HEK293 total RNA with head-to-tail specific probes for circRNAs. PAR-CUPcontrols GAPDH, linear control. d, e, circRNAs are at least 10-fold more RNase R resistant than GAPDH mRNA (d) and stable after 24 h transcription block (e)(qPCR; error bars indicate standard deviation). cr. All of these were at least 10-fold more resistant than GAPdH 5 ig. 2d and Supplementary Fig. 2a). We reasoned that circRNAs 8-14 should generally turn over more slowly than mRNAS. Indeed, we 5-4 found that 24 h after blocking transcription circRNAs were highly stable, exceeding the stability of the housekeeping gene GAPDH' (Fig. 2e and Supplementary Fig. 2b). We also validated 3 out of 3 tested mouse circRNAs with human orthologues in mouse brains HEK293 expression relative to 18S(%) (Supplementary Fig. 2c). In C elegans 15 out of 20(75%)of the pre dictions from gametes and early embryos were validated in a mixed stage sample( Supplementary Fig 2d and Supplementary Table 3) SSCR circRNA CDRlas is densely bound by AGo APDH CDRlas Stable transcripts with many miRNA-binding sites could function as miRNA sponges. We intersected our catalogue of circRNAs with transcript annotations, assuming that introns would not occur in mature circRNAs (as observed for 3 out of 3 tested circRNAs, Supplementary Fig. 2e). We screened for occurrences of conserved miRNA family seed matches(Methods). When counting repetitions of conserved matches to the same miRNA family, circRNAs were Table 1 Summary of the validation experiments Figure 3 The circRNA CDRlas is bound by the miRNA effector protein Sample Validation experiment AGO, and is cytoplasmic. a, CDRlas is densely bound by AGo (red)but not by unrelated proteins(black). Blue boxes indicate miR-7 seed matches. nt, Human(HEK293) 19of23 nudeotides. b, c, miR-7 sites display reduced nucleotide variability across 32 Expression >3% vinculin 120f21 vertebrate genomes( b)and high base-pairing probability within seed matches c).d, CDRlas RNA is cytoplasmic and disperse(white spots; single-mole RNA FISH; maximum intensity merges of Z-stacks). siSCR, positive; siRNAl Mouse(adult brain Head-to-tail splicing 3 of 3 negative control. Blue, nuclei(DAPI); scale bar, 5 um(see also Supr Fig 10 for uncropped images).e, Northern blotting detects circular but not Expression>1%β~ actin linear CDRlas in HEK293 RNA. Total, HEK293 RNA; circular, head-to-tail 15of20 probe; circ+lin, probe within splice sites; IVT lin, in vitro transcribed, linear CDRlas RNA f, Circular CDRlas is highly expressed (qPCR, error bars pression>1%ei·3d 120f15 indicate standard deviation). g, CDRlas. Blue, seed matches; dark red, AGO PAR-CLIP reads; bright red, crosslinked nucleotide conversions. 00 MONTH 2013 VOL 000 NATURE 3 @2013 Macmillan Publishers Limited. All rights reserved

qPCR. All of these were at least 10-fold more resistant than GAPDH (Fig. 2d and Supplementary Fig. 2a). We reasoned that circRNAs should generally turn over more slowly than mRNAs. Indeed, we found that 24 h after blocking transcription circRNAs were highly stable, exceeding the stability of the housekeeping gene GAPDH31 (Fig. 2e and Supplementary Fig. 2b). We also validated 3 out of 3 tested mouse circRNAs with human orthologues in mouse brains (Supplementary Fig. 2c). In C. elegans 15 out of 20 (75%) of the pre￾dictions from gametes and early embryos were validated in a mixed stage sample (Supplementary Fig. 2d and Supplementary Table 3). circRNA CDR1as is densely bound by AGO Stable transcripts with many miRNA-binding sites could function as miRNA sponges. We intersected our catalogue of circRNAs with transcript annotations, assuming that introns would not occur in mature circRNAs (as observed for 3 out of 3 tested circRNAs, Supplementary Fig. 2e). We screened for occurrences of conserved miRNA family seed matches (Methods). When counting repetitions of conserved matches to the same miRNA family, circRNAs were significantly enriched compared to coding sequences (P , 2.963 10222, Mann–Whitney U-test, n 5 3,873) or 39 UTR sequences (P , 2.76 3 10221, Mann–Whitney U-test, n 5 3,182) (Supplementary Fig. 3a, b). As an extreme case, we discovered that the known human circRNA CDR1as (ref. 9) harboured dozens of conserved miR-7 seed matches. To test whether CDR1as is bound by miRNAs, we analysed bio￾chemical, transcriptome-wide binding-site data for the miRNA effector AGO proteins. We performed four independent PAR-CLIP (photoactivatable-ribonucleoside-enhanced crosslinking and immu￾noprecipitation) experiments for human AGO (Methods) and ana￾lysed them together with published, lower-depth data32. PAR-CLIP32–34 is based on ultraviolet crosslinking of RNA to protein and subsequent sequencing of RNA bound to a RBP of interest. The ,1.5-kilobase (kb) CDR1as locus stood out in density and number of AGO PAR-CLIP reads (Fig. 3a), whereas nine combined PAR-CLIP libraries for other RBPs gave virtually no signal. Of note, there is no PAR-CLIP read mapping to the sense coding transcript of the CDR1 gene, which was originally identified as a target of autoantibodies from patients with paraneoplastic cerebellar degeneration35. Sequence analysis across 32 vertebrate species revealed that miR-7 is the only animal miRNA with conserved seed matches that can explain the AGO binding along the CDR1as transcript (Methods). Human CDR1as harbours 74 miR-7 seed matches of which 63 are Table 1 | Summary of the validation experiments Sample Validation experiment Validation success Human (HEK293) Head-to-tail splicing 19 of 23 Circularity 21 of 21 Expression .3% vinculin 12 of 21 Expression specificity (leukocyte specific) 5 of 7 Mouse (adult brain) Head-to-tail splicing 3 of 3 Circularity 3 of 3 Expression .1% b-actin 2 of 3 C. elegans Head-to-tail splicing 15 of 20 Circularity 13 of 13 Expression .1% eif-3.d 12 of 15 Most experimentally tested circRNAs are validated. 74 miR-7 seed matches AGO PAR-CLIP reads CDR1 antisense (1,485 nt) 5′ 3′ 5′ 3′ Splice junction miR-7 seed miR-7 (nt) 5′ to 3′ Probability 1.0 0.5 UGGAAGA CUAGUGAUUUUGUUGU Base pairing Distance from seed match (nt) –entropy (bits) –1 –1.8 –2.2 –8 GUCUUCCA +8 Seed match conservation miR-7 seed match PAR-CLIP controls (QKI, PUM2, ELAVL1) HEK293 expression relative to 18S (%) 1.0 0.2 0.6 Circular GAPDH CDR1as a b c d e f VCL g RNA gel 28S 18S 5S Circular circ+lin Total IVT lin. Total IVT lin. Total IVT lin. siSCR siRNA1 Mock Figure 3 | The circRNA CDR1as is bound by the miRNA effector protein AGO, and is cytoplasmic. a, CDR1as is densely bound by AGO (red) but not by unrelated proteins (black). Blue boxes indicate miR-7 seed matches. nt, nucleotides. b, c, miR-7 sites display reduced nucleotide variability across 32 vertebrate genomes (b) and high base-pairing probability within seed matches (c). d, CDR1as RNA is cytoplasmic and disperse (white spots; single-molecule RNA FISH; maximum intensity merges of Z-stacks). siSCR, positive; siRNA1, negative control. Blue, nuclei (DAPI); scale bar, 5 mm (see also Supplementary Fig. 10 for uncropped images). e, Northern blotting detects circular but not linear CDR1as in HEK293 RNA. Total, HEK293 RNA; circular, head-to-tail probe; circ1lin, probe within splice sites; IVT lin., in vitro transcribed, linear CDR1as RNA. f, Circular CDR1as is highly expressed (qPCR, error bars indicate standard deviation). g, CDR1as. Blue, seed matches; dark red, AGO PAR-CLIP reads; bright red, crosslinked nucleotide conversions. 5S 18S 28S RNase R exonuclease Agarose gel GAPDH CDR1as hsa-circRNAs 2 3 16 – + Divergent Convergent hsa-circRNA2 hsa-circRNA3 hsa-circRNA16 GAPDH cDNA gDNA gDNA gDNA gDNA 100 200 300 cDNA cDNA cDNA Sanger sequencing ... ... GAPDH p16 Mock 24 h ActD 0.0 0.5 1.0 Rel. expression CDR1as hsa-circRNAs 2 3 6 9 16 ZRANB1 exon1 hsa-circRNA2 ... ... a b c d GAPDH VCL CDR1as hsa-circRNAs 2 3 6 9 16 0.0 0.5 1.0 Rel. expression Mock RNase R e AG GT – + – + – + – + – + ( ) ( ) Figure 2 | CircRNAs are stable transcripts with robust expression. a, Human (hsa) ZRANB1 circRNA exemplifies the validation strategy. Convergent (divergent) primers detect total (circular) RNAs. Sanger sequencing confirms head-to-tail splicing. b, Divergent primers amplify circRNAs in cDNA but not genomic DNA (gDNA). GAPDH, linear control, size marker in base pairs. c, Northern blots of mock (2) and RNase R (1) treated HEK293 total RNA with head-to-tail specific probes for circRNAs. GAPDH, linear control. d, e, circRNAs are at least 10-fold more RNase R resistant than GAPDH mRNA (d) and stable after 24 h transcription block (e) (qPCR; error bars indicate standard deviation). ARTICLE RESEARCH 00 MONTH 2013 | VOL 000 | NATURE | 3 ©2013 Macmillan Publishers Limited. All rights reserved

RESEARCH ARTICLE conserved in at least one other species(Supplementary Fig. 4). CDR1as and miR-7 in mouse tissues Interspaced sequences were less conserved, indicating that miR-7 32gCERs analysis of predicted circRNA-mirna duplexes(Methods)showed 3 reduced base-pairing of miR-7 beyond the seed( Fig 3c). None of the 2 1, 500 miR-7 complementary sites across 32 vertebrate sequences 8 was complementary beyond position 12 of miR-7(only three could form an 11-nucleotide duplex)(Supplementary Table 4). Slicing by …x mammalian Argonaute requires complementarity of positions 10 and I and depends on extended complementarity beyond position 12 -124 control (ref. 36). Thus, CDRlas seems optimized to be densely bound but not liced by miR-7 Single-molecule imaging(Methods) revealed disperse and most ? ytoplasmic CDRlas expression(HEK293 cells), consistent with miRNA sponge function (Fig. 3d and Supplementary Table 5). Figure 4 CDRlas and miR-7 have ove CDRlas circularization was assayed by northern blotting(Fig. 3e). neuronal tissues. a, Among mouse tissues and MIN6 cells(qPCR,relative to Nicking experiments confirmed that CDRlas circRNA can be linea- cerebral cortex expression; error bars indicate standard deviations; see rized and degraded(Supplementary Fig. 5a). In RNA from HEK293 Supplementary Fig 9a for miR-122 control) neuronal tissues co-express miR-7 cells, circularized but no additional linear CDRlas was detected and CDRlas. b, In situ staining of CDRlas and miR-7 in mouse embryo brain (Supplementary Fig. 5b). Circular expression levels were quantified E13.5(U6 and miR-124, positive control; scrambled probe, negative control) by qPCR with divergent primers calibrated by standard curves (Supplementary Table 6). CDRlas was highly expressed(-15%to 20% of GAPDH expression, Fig 3f). Estimating GAPDH mRNA unknown consequences. This problem is circumvented when using zeb- copy number from HEK293 RNA-seq data(-1, 400 molecules per rafish( Danio rerio)as an animal modeL. According to our bioinformatic cell, data not shown)suggests that CDRlas may bind up to -20,000 analyses (not shown )zebrafish has lost the cdrl locus, whereas miR-7 is miR-7 molecules per cell( Fig. 3g) conserved and highly expressed in the embryonic l brain". Thus, we can If CDRlas functions as a miR-7 sponge, its destruction could trigger test whether miR-7 has a loss-of-function phenotype and if this pheno- downregulation of miR-7 targets. We knocked down CDRlas in type can be induced by introduction of mammalian CDRlas rNA We HEK293 cells and monitored of published miR-7 targets injected morpholinos to knock down mature mir-7 expression in zebra by q PCR with externally spiked-in standards(Methods and Supplemen- fish embryos(Methods). At a dose of 9 ng of miR-7 morpholino, the ary Fig 5c, d). All eight miR-7 targets assayed, but also housekeeping embryos did not show overall morphological defects but reproducibly genes, were downregulated Nanostring technology ?7 additionally indi- and in two independent genetic backgrounds(Supplementary Fig 6a-c) ated downregulation of many genes(data not shown). Furthermore, developed brain defects(Fig, 5a, b). In particular, -70% showed a con- stable loss of CDRlas expression by virally delivered small hairpin sistent and clear reduction in midbrain size, and an additional-5% of RNAs led to significantly reduced migration in an in vitro wound clo- animals had almost completely lost their midbrains. Of note, the tel- sure assay(Methods, Supplementary Fig, 5e, f and Supplementary encephalon at the anterior tip of the brain was not affected in size. Brain Table 7). Thus, knockdown of CDRlas affects HEK293 cells, but we volumes were also measured based on confocal three-dimensional stacks could not delineate miR-7-specificeffects, potentially because ofindirect (Fig. 5c and Supplementary Fig. 7). Reduction of the midbrain size or miR-7-independent CDRlas function(see below). correlated with miR-7 inhibition in the respective animals(Supplemen tary Fig. 6d). These data provide evidence that miR-7 loss-of-function Co-expression of miR-7 and CDRlas in brain causes a specific reduction of midbrain size If CDRlas indeed interacts with miR-7, both must be co-expressed. To test whether CDRIas can function as a miR-7 sponge in vivo,we miR-7 is highly expressed in neuronal tissues, pancreas and pituit injected embryos with plasmid DNA that expressed a linear version of glands. apart from HEK293 cells, a cell line probably derived from the full-length human CDRlas sequence(Supplementary Fig 6e, f)or neuronal precursors in embryonic kidney 9, we quantified miR-7 and a plasmid provided by the Kjems laboratory that can produce circular CDRlas expression across mouse tissues and pancreatic- island- CDRIas in human cells(Fig. 5d, e). q PCR analysis detected circular derived MIN6 cells(Methods and Fig 4a). CDRlas and miR-7 were RNA in zebrafish embryos injected with the latter plasmid (Sup both highly expressed in brain tissues, but CDRlas was expressed at plementary Fig 8), which reproducibly and in independent genetic low levels or absent in non-neuronal tissues, including tissues with backgrounds lead to reduced midbrain sizes(Fig. 5g, h). Similarly very high miR-7 expression. qPCR suggested that CDRlas is exclu- animals injected with in vitro-transcribed partial mouse CDRlas sively circular in adult and embryonic mouse brain( Supplemen RNA, but not with RNA from the other strand, showed significant Fig. 5g, h). Thus, CDRlas and miR-7 seem to interact specifically midbrain reduction(Supplementary Fig 6g-i). Thus, the phenotype in neuronal tissues. Indeed, when assaying CDRlas and miR-7 in is probably caused by CDRlas rna and not by an unspecific effect of mouse brains by in situ hybridizations(Methods), we observed RNA or DNA injection. These results provide evidence that human/ cific, similar, but not identical, expression patterns in the brain of mouse CDRlas transcripts are biologically active in vivo and impai mid-gestation(embryonic day 13.5(E13.5)embryos( Fig 4b). Speci- brain development similarly to miR-7 inhibition. The midbrain fically, CDRlas and miR-7 were highly co-expressed in areas of the reduction could be partially rescued by inje developing midbrain(mesencephalon)*04. Thus, CDRlas is highly (Fig. 5f, g), arguing that the biological effect of CDRlas expression expressed, stable, cytoplasmic, not detectable as a linear RNA and is caused at least in part by interaction of CDRlas with miR-7 shares expression domains with miR-7. Together with extensive miR-7 binding within CDRlas, CDRlas has hallmarks of a potent Discussion We have shown that animal genomes express thousands of circRNAs from diverse genomic locations(for example, from coding and non- Effects of miR-7 and CDRlas in zebrafish coding exons, intergenic regions or transcripts antisense to 5 system. However, a knockout would also affect CDRI protein, with specific manner h plex tissue, cell-type- or developmental-stage- It would be informative to knock out CDRlas in an animal model 3 UTRs)in a cor rovided evidence that cDrlas can act as a 4I NATURE I VOL 00000 MONTH 2013 @2013 Macmillan Publishers Limited. All rights reserved

conserved in at least one other species (Supplementary Fig. 4). Interspaced sequences were less conserved, indicating that miR-7 binding sites are probably functional (Fig. 3b). Secondary structure analysis of predicted circRNA–miRNA duplexes (Methods) showed reduced base-pairing of miR-7 beyond the seed (Fig. 3c). None of the ,1,500 miR-7 complementary sites across 32 vertebrate sequences was complementary beyond position 12 of miR-7 (only three could form an 11-nucleotide duplex) (Supplementary Table 4). Slicing by mammalian Argonaute requires complementarity of positions 10 and 11 and depends on extended complementarity beyond position 12 (ref. 36). Thus, CDR1as seems optimized to be densely bound but not sliced by miR-7. Single-molecule imaging (Methods) revealed disperse and mostly cytoplasmic CDR1as expression (HEK293 cells), consistent with miRNA sponge function (Fig. 3d and Supplementary Table 5). CDR1as circularization was assayed by northern blotting (Fig. 3e). Nicking experiments confirmed that CDR1as circRNA can be linea￾rized and degraded (Supplementary Fig. 5a). In RNA from HEK293 cells, circularized but no additional linear CDR1as was detected (Supplementary Fig. 5b). Circular expression levels were quantified by qPCR with divergent primers calibrated by standard curves (Supplementary Table 6). CDR1as was highly expressed (,15% to ,20% of GAPDH expression, Fig. 3f). Estimating GAPDH mRNA copy number from HEK293 RNA-seq data (,1,400 molecules per cell, data not shown) suggests that CDR1as may bind up to ,20,000 miR-7 molecules per cell (Fig. 3g). If CDR1as functions as a miR-7 sponge, its destruction could trigger downregulation of miR-7 targets. We knocked down CDR1as in HEK293 cells and monitored expression of published miR-7 targets by qPCR with externally spiked-in standards (Methods and Supplemen￾tary Fig. 5c, d). All eight miR-7 targets assayed, but also housekeeping genes, were downregulated. Nanostring technology37 additionally indi￾cated downregulation of many genes (data not shown). Furthermore, stable loss of CDR1as expression by virally delivered small hairpin RNAs led to significantly reduced migration in an in vitro wound clo￾sure assay (Methods, Supplementary Fig. 5e, f and Supplementary Table 7). Thus, knockdown of CDR1as affects HEK293 cells, but we could not delineate miR-7-specific effects, potentially because of indirect or miR-7-independent CDR1as function (see below). Co-expression of miR-7 and CDR1as in brain If CDR1as indeed interacts with miR-7, both must be co-expressed. miR-7 is highly expressed in neuronal tissues, pancreas and pituitary gland38. Apart from HEK293 cells, a cell line probably derived from neuronal precursors in embryonic kidney39, we quantified miR-7 and CDR1as expression across mouse tissues and pancreatic-island￾derived MIN6 cells (Methods and Fig. 4a). CDR1as and miR-7 were both highly expressed in brain tissues, but CDR1as was expressed at low levels or absent in non-neuronal tissues, including tissues with very high miR-7 expression. qPCR suggested that CDR1as is exclu￾sively circular in adult and embryonic mouse brain (Supplementary Fig. 5g, h). Thus, CDR1as and miR-7 seem to interact specifically in neuronal tissues. Indeed, when assaying CDR1as and miR-7 in mouse brains by in situ hybridizations (Methods), we observed spe￾cific, similar, but not identical, expression patterns in the brain of mid-gestation (embryonic day 13.5 (E13.5)) embryos (Fig. 4b). Speci￾fically, CDR1as and miR-7 were highly co-expressed in areas of the developing midbrain (mesencephalon)40,41. Thus, CDR1as is highly expressed, stable, cytoplasmic, not detectable as a linear RNA and shares expression domains with miR-7. Together with extensive miR-7 binding within CDR1as, CDR1as has hallmarks of a potent circular miR-7 sponge in neuronal tissues. Effects of miR-7 and CDR1as in zebrafish It would be informative to knock out CDR1as in an animal model system. However, a knockout would also affect CDR1 protein, with unknown consequences. This problem is circumvented when using zeb￾rafish (Danio rerio) as an animal model. According to our bioinformatic analyses (not shown) zebrafish has lost the cdr1 locus, whereas miR-7 is conserved and highly expressed in the embryonic brain42. Thus, we can test whether miR-7 has a loss-of-function phenotype and if this pheno￾type can be induced by introduction of mammalian CDR1as RNA. We injected morpholinos to knock down mature miR-7 expression in zebra￾fish embryos (Methods). At a dose of 9 ng of miR-7 morpholino, the embryos did not show overall morphological defects but reproducibly, and in twoindependent genetic backgrounds (Supplementary Fig. 6a–c), developed brain defects (Fig. 5a, b). In particular, ,70% showed a con￾sistent and clear reduction in midbrain size, and an additional ,5% of animals had almost completely lost their midbrains. Of note, the tel￾encephalon at the anterior tip of the brain was not affected in size. Brain volumes were also measured based on confocal three-dimensional stacks (Fig. 5c and Supplementary Fig. 7). Reduction of the midbrain size correlated with miR-7 inhibition in the respective animals (Supplemen￾tary Fig. 6d). These data provide evidence that miR-7 loss-of-function causes a specific reduction of midbrain size. To test whether CDR1as can function as a miR-7 sponge in vivo, we injected embryos with plasmid DNA that expressed a linear version of the full-length human CDR1as sequence (Supplementary Fig. 6e, f) or a plasmid provided by the Kjems laboratory that can produce circular CDR1as in human cells (Fig. 5d, e). qPCR analysis detected circular RNA in zebrafish embryos injected with the latter plasmid (Sup￾plementary Fig. 8), which reproducibly and in independent genetic backgrounds lead to reduced midbrain sizes (Fig. 5g, h). Similarly, animals injected with in vitro-transcribed partial mouse CDR1as RNA, but not with RNA from the other strand, showed significant midbrain reduction (Supplementary Fig. 6g–i). Thus, the phenotype is probably caused by CDR1as RNA and not by an unspecific effect of RNA or DNA injection. These results provide evidence that human/ mouse CDR1as transcripts are biologically active in vivo and impair brain development similarly to miR-7 inhibition. The midbrain reduction could be partially rescued by injecting miR-7 precursor (Fig. 5f, g), arguing that the biological effect of CDR1as expression is caused at least in part by interaction of CDR1as with miR-7. Discussion We have shown that animal genomes express thousands of circRNAs from diverse genomic locations (for example, from coding and non￾coding exons, intergenic regions or transcripts antisense to 59 and 39 UTRs) in a complex tissue-, cell-type- or developmental-stage￾specific manner. We provided evidence that CDR1as can act as a CDR1as and miR-7 expression in mouse tissues CDR1as expression CDR1as miR-7 400 miR-7 expression 0 1 2 1 10 600 Hippocampus Cer. cortex Cerebellum Forebrain Midbrain Kidney Lung Skel. muscle Spleen Pancreas Liver Colon Heart Pituitary gland MIN6 cells 5 miR-7 U6 control miR-124 control CDR1as Scrambled a b Figure 4 | CDR1as and miR-7 have overlapping and specific expression in neuronal tissues. a, Among mouse tissues and MIN6 cells (qPCR, relative to cerebral cortex expression; error bars indicate standard deviations; see Supplementary Fig. 9a for miR-122 control) neuronal tissues co-express miR-7 and CDR1as. b, In situ staining of CDR1as and miR-7 in mouse embryo brain E13.5 (U6 and miR-124, positive control; scrambled probe, negative control). Scale bar, 1 mm. RESEARCH ARTICLE 4 | NATURE | VOL 000 | 00 MONTH 2013 ©2013 Macmillan Publishers Limited. All rights reserved

ARTICLE RESEARCH Control Mo 15 ng MO g CDRlas. Thus, CDRlas may function to transport miR-7 to subcel- lular locations, where miR-671 could trigger release of its cargo. Known functions of miR-7 targets such as PAKI and FAKI support these peculation The phenotype induced by CDRlas expression in zebrafish was only partially rescued by expressing miR-7, indicating that CDRlas could have functions beyond sequestering miR-7. This idea is sup- ported by in situ hybridization in mouse adult hippocampus(Su ed Me plementary Fig 9b)where areas staining for CDRlas but not miR- were observed What could be additional functions of circrnas IC egfp) yond acting as sponges? As a single-stranded RNA, CDRlas could Circular CDR1as Circular CDRlas +miR-7 for example, bind in trans 3 UTRs of target mRNAs to regul expression. It is even possible that miR-7 binds CDRlas to these trans-acting activities. Alternatively, CDRlas could be in the assembly of larger complexes of RNA or protein, similar to other low-complexity molecules How many other circRNAs exist? In this study, we identifie ximately 2,000 human, 1,900 mouse and 700 nematode from sequencing data, and our validation experiments most of the 50 tested circRNAS. However, we analysed only a few tissues/developmental stages with stringent cutoffs. Thus, the true number of circRNAs is almost certainly much larger. Although 图 h且x星 CDRlas is an extreme cas circRNAs have conserved seed matches. For example, circRNA from the SRY locus has seed sites for murine miRNAs. Therefore, circRNAs probably compete with 圖 22 other RNAs for miRNA binding. Sequence analyses indicated that oding exons serve additional, presumably regulatory functions when expressed within circRNAs, whereas intergenic or intronic circRNAs generally showed only weak conservation. Because we detected thou Mo (ng) CDR1 larization of exons is easy to evolve and may provide a mechanism Figure 5 In zebrafish, knockdown of miR-7 or expression of CDRlas for rapid evolution of stably and well expressed regulatory RNAs Of uses midbrain defects. a, b, Neuronal ter(Tg(huc egp)embryos(top, note, we detected multiple seed matches for viral miRNAs within light microscopy)48 h post fertilization( bottom, representative confoc human circRNAs(not shown). However, there is no reason to think stack projections; blue dashed line, telencephalon(TC)(control); yellow that circRNAs function predominantly to bind miRNAs. As known dashed line, midbrain(MB). Embryos after injection of 9 ng miR-7 in bacteria, the decoy mechanism underlying miRNA sponges could morpholino(MO)(b)display a reduction in midbrain size. Panel a shows a be important also for RBPs"647. Similarly, circRNAs could function to store, sort, or localize RBPs. In summary, our data suggest that vector encoding human circular CDRlas. f, Rescue experiment with miR-7 circRNAs form a class of post-transcriptional regulators which com- dimensional volumetric reconstructions. d, Empty vector control e, Expr precursor g, Phenotype penetrance(% of embryos, miR-7 MO, n=135: pete with other RNAs for binding by miRNAs and RBPs and may 91; linear CDRlas, n=258; circular generally function in modulating the local free concentration of RBPs, CDRlas, n=153; circular CDRlas plus miR-7 precursor, n=217). Phenotype RNAS, or their binding sites distribution derived from at least three independent experiments. Scale bar, Note added in proof: While this paper was under review, circular 0.1 mm. *P<ool: ***p< o0o1 in Students t-test for normal midbrain RNAs in fibroblasts were described reduced midbrain(see also Supplementary Fig. 6).h, Phenotype quantification (Methods). Error bars indicate standard deviation n= 3 per group. METHODS SUMMARY CDRlas is densely bound by miRNA effector molecules;(2)CDRlas f ode e al pipeline for predicting circRNAs from ribominus sequencing post-transcriptional regulator by binding miR-7 in brain tissues:(1) data. a detailed description of the computational methods is given in the is expressed highly, stably and mostly cytoplasmic: (4)CDRlas and T-REx (Life Technologies) were caultured-following standard protocols. Tran. entre Biotechnologies)treatment (3U ug )was performed on total RNA (5)human/mouse CDRlas is circularized in vivo and is not detectable (5 ug)at 37C for 15 min qPCR primers are listed in Supplementary Table 8 as a linear molecule;(6) human/mouse CDRlas sequences, when Single-molecule RNA fluorescence in situ hybridization(smRNA FISH) injected into zebrafish, and miR-7 knock down have similar pheno- Stellaris Oligonucleotide probes complementary to CDRlas were designed using types in brain. While zebrafish circularization of human CDRlas may the Stellaris Probe Designer(Biosearch Technologies). Probe pools were obtained be incomplete, the midbrain phenotype was stronger compared to from Bio Cat GmbH as conjugates coupled to Quasar 670 Probes were hybridized expressing linear CDRlas RNA that lacks circularization splice sites. at 125 nM at 37 C Images were acquired on an inverted Nikon Ti microscope Although the two DNA plasmids used carry identical pre s and Mouse strains and in situ hybridization. In situ hybridization(ISH)was per- that the difference in midbrain phenotype strength may be explained using locked nucleic acid (LNA) probes or RNAs obtained by in vitro transcrip- by other factors. However, because of the observed extreme stability zebrafish methods. Tg(hu C egfp )and Tg(Xia. Tubb: ds RED)transgenic zebrafish of CDRlas and circRNAs in general, our data argue that circrNAs lines were used 9 0. Morpholino antisense oligomers were injected into the yolk of an be used as potent inhibitors of miRNAs or RBPs. Future studies single-cell-stage embryos. Furthermore, two pCS2+ plasmids coding for full nould elucidate how CDRlas can be converted into a linear mole- length linear CDRlas or CDRlas plus upstream and downstream sequence that ule and targeted for degradation. miR-671 can trigger destruction of can express circular CDRlas in human cells(courtesy of the Kjems laboratory) 00 MONTH 2013 VOL 000 NATURE I @2013 Macmillan Publishers Limited. All rights reserved

post-transcriptional regulator by binding miR-7 in brain tissues: (1) CDR1as is densely bound by miRNA effector molecules; (2) CDR1as harbours 74 miR-7 seed matches, often deeply conserved; (3) CDR1as is expressed highly, stably and mostly cytoplasmic; (4) CDR1as and miR-7 share specific expression domains in mouse embryonic brain; (5) human/mouse CDR1as is circularized in vivo and is not detectable as a linear molecule; (6) human/mouse CDR1as sequences, when injected into zebrafish, and miR-7 knock down have similar pheno￾types in brain. While zebrafish circularization of human CDR1as may be incomplete, the midbrain phenotype was stronger compared to expressing linear CDR1as RNA that lacks circularization splice sites. Although the two DNA plasmids used carry identical promoters and were injected in equal concentrations, we cannot rule out the possibility that the difference in midbrain phenotype strength may be explained by other factors. However, because of the observed extreme stability of CDR1as and circRNAs in general, our data argue that circRNAs can be used as potent inhibitors of miRNAs or RBPs. Future studies should elucidate how CDR1as can be converted into a linear mole￾cule and targeted for degradation. miR-671 can trigger destruction of CDR1as9 . Thus, CDR1as may function to transport miR-7 to subcel￾lular locations, where miR-671 could trigger release of its cargo. Known functions of miR-7 targets such as PAK1 and FAK1 support these speculations43,44. The phenotype induced by CDR1as expression in zebrafish was only partially rescued by expressing miR-7, indicating that CDR1as could have functions beyond sequestering miR-7. This idea is sup￾ported by in situ hybridization in mouse adult hippocampus (Sup￾plementary Fig. 9b) where areas staining for CDR1as but not miR-7 were observed. What could be additional functions of circRNAs beyond acting as sponges? As a single-stranded RNA, CDR1as could, for example, bind in trans 39 UTRs of target mRNAs to regulate their expression. It is even possible that miR-7 binds CDR1as to silence these trans-acting activities. Alternatively, CDR1as could be involved in the assembly of larger complexes of RNA or protein, perhaps similar to other low-complexity molecules45. How many other circRNAs exist? In this study, we identified appro￾ximately 2,000 human, 1,900 mouse and 700 nematode circRNAs from sequencing data, and our validation experiments confirmed most of the 50 tested circRNAs. However, we analysed only a few tissues/developmental stages with stringent cutoffs. Thus, the true number of circRNAs is almost certainly much larger. Although CDR1as is an extreme case, many circRNAs have conserved seed matches. For example, circRNA from the SRY locus6 has seed sites for murine miRNAs. Therefore, circRNAs probably compete with other RNAs for miRNA binding. Sequence analyses indicated that coding exons serve additional, presumably regulatory functions when expressed within circRNAs, whereas intergenic or intronic circRNAs generally showed only weak conservation. Because we detected thou￾sands of circRNAs, it is appealing to speculate that occasional circu￾larization of exons is easy to evolve and may provide a mechanism for rapid evolution of stably and well expressed regulatory RNAs. Of note, we detected multiple seed matches for viral miRNAs within human circRNAs (not shown). However, there is no reason to think that circRNAs function predominantly to bind miRNAs. As known in bacteria, the decoy mechanism underlying miRNA sponges could be important also for RBPs46,47. Similarly, circRNAs could function to store, sort, or localize RBPs. In summary, our data suggest that circRNAs form a class of post-transcriptional regulators which com￾pete with other RNAs for binding by miRNAs and RBPs and may generally function in modulating the local free concentration of RBPs, RNAs, or their binding sites. Note added in proof: While this paper was under review, circular RNAs in fibroblasts were described51. METHODS SUMMARY Computational pipeline for predicting circRNAs from ribominus sequencing data. A detailed description of the computational methods is given in the Methods. Cell culture and treatments. HEK293, HEK293TN and HEK293 Flp-In 293 T-REx (Life Technologies) were cultured following standard protocols. Tran￾scription was blocked by adding 2 mg ml21 actinomycin D (Sigma). RNase R (Epicentre Biotechnologies) treatment (3 U mg21 ) was performed on total RNA (5 mg) at 37 uC for 15 min. qPCR primers are listed in Supplementary Table 8. Single-molecule RNA fluorescence in situ hybridization (smRNA FISH). Stellaris Oligonucleotide probes complementary to CDR1as were designed using the Stellaris Probe Designer (Biosearch Technologies). Probe pools were obtained from BioCat GmbH as conjugates coupled to Quasar 670. Probes were hybridized at 125 nM at 37 uC. Images were acquired on an inverted Nikon Ti microscope. Mouse strains and in situ hybridization. In situ hybridization (ISH) was per￾formed on paraffin tissue sections from B6129SF1/J wild-type mice as described48 using locked nucleic acid (LNA) probes or RNAs obtained by in vitro transcrip￾tion on PCR products. Zebrafish methods. Tg(huC:egfp) and Tg(Xia.Tubb:dsRED) transgenic zebrafish lines were used49,50. Morpholino antisense oligomers were injected into the yolk of single-cell-stage embryos. Furthermore, two pCS21 plasmids coding for full￾length linear CDR1as or CDR1as plus upstream and downstream sequence that can express circular CDR1as in human cells (courtesy of the Kjems laboratory) Empty vector Circular CDR1as Circular CDR1as + miR-7 a Control MO 15 ng b c d e MB TC f g 3D reconstruction Control MO miR-7 MO Uninjected Linear CDR1as Circular CDR1as Volume (×106 μm3) 1 2 ** *** ** MB h TC ** Phenotype (%) 0 Empty vector Linear Circular Circular + miR-7 miR-7 (9 ng) Control (15 ng) Uninjected ****** ** ** 100 *** 50 Normal Reduced MB No MBmiR-7 (15 ng) Normal MB Reduced MB miR-7 MO 9 ng MO (ng) CDR1as 3 Tg(huC:egfp) f Circular CDR1as + miR-7 TC Figure 5 | In zebrafish, knockdown of miR-7 or expression of CDR1as causes midbrain defects. a, b, Neuronal reporter (Tg(huC:egfp)) embryos (top, light microscopy) 48 h post fertilization (bottom, representative confocal z-stack projections; blue dashed line, telencephalon (TC) (control); yellow dashed line, midbrain (MB)). Embryos after injection of 9 ng miR-7 morpholino (MO) (b) display a reduction in midbrain size. Panel a shows a representative embryo injected with 15 ng control morpholino. c, Three￾dimensional volumetric reconstructions. d, Empty vector control. e, Expression vector encoding human circular CDR1as. f, Rescue experiment with miR-7 precursor. g, Phenotype penetrance (% of embryos, miR-7 MO, n 5 135; uninjected, n 5 83; empty vector, n 5 91; linear CDR1as, n 5 258; circular CDR1as, n 5 153; circular CDR1as plus miR-7 precursor, n 5 217). Phenotype distribution derived from at least three independent experiments. Scale bar, 0.1 mm. **P , 0.01; ***P , 0.001 in Students t-test for normal midbrain, reduced midbrain (see also Supplementary Fig. 6). h, Phenotype quantification (Methods). Error bars indicate standard deviation n 5 3 per group. ARTICLE RESEARCH 00 MONTH 2013 | VOL 000 | NATURE | 5 ©2013 Macmillan Publishers Limited. All rights reserved