Abasic pivot substitution harnesses target specificity of RNA interference

Gene silencing via RNA interference inadvertently represses hundreds of off-target transcripts. Because small interfering RNAs (siRNAs) can function as microRNAs, avoiding miRNA-like off-target repression is a major challenge. Functional miRNA–target interactions are known to pre-require transitional nucleation, base pairs from position 2 to the pivot (position 6). Here, by substituting nucleotide in pivot with abasic spacers, which prevent base pairing and alleviate steric hindrance, we eliminate miRNA-like off-target repression while preserving on-target activity at ∼80–100%. Specifically, miR-124 containing dSpacer pivot substitution (6pi) loses seed-mediated transcriptome-wide target interactions, repression activity and biological function, whereas other conventional modifications are ineffective. Application of 6pi allows PCSK9 siRNA to efficiently lower plasma cholesterol concentration in vivo, and abolish potentially deleterious off-target phenotypes. The smallest spacer, C3, also shows the same improvement in target specificity. Abasic pivot substitution serves as a general means to harness the specificity of siRNA experiments and therapeutic applications.

T he RNA-induced silencing complex (RISC) is responsible for inducing gene silencing caused by RNA interference (RNAi), guided by B21 to 23 nucleotides (nt) duplexes of RNAs in a sequence-dependent manner 1 . As regulators of biological function, microRNAs (miRNAs) are endogenously produced to form such structures, of which one strand termed 'guide strand' directs the RISC to post-transcriptionally repress target genes by altering mRNA stability and/or translation 2,3 . Functional miRNA-target interactions generally require partial sequence complementarity, majorly pairing as few as 6 nt matches within the seed region (positions 2-8; ref. 4). To utilize the RISC to selectively silence a desired target gene, small interfering RNAs (siRNAs) are typically designed and synthesized as duplexes, in which the guide strand is perfectly complementary to the target mRNA. When siRNAs are introduced into the cell 5 , the guide strand triggers cleavage of the intended target mRNA by loading into Argonaute (Ago2, also known as Eif2c2), the core catalytic component of the RISC 6 . However, a critical caveat in using siRNAs is that any B21 to 23 nt RNA incorporated into Argonaute (Ago) can also function as miRNA, by the mechanism via which miRNAs recognize target mRNAs 7,8 . Thus, the use of siRNAs always results in the repression of hundreds of off-target transcripts, thereby potentially leading to unintended phenotypes.
A number of studies have demonstrated that the off-target effects of siRNAs are widespread throughout transcripts 8,9 and seed-centric in terms of recognition 10,11 as shown with miRNAs 4 . Compared with cleavage-mediated on-target silencing, off-target activity is relatively modest at an individual transcript level 9,10 , and often considered phenotypically irrelevant 8 . Nevertheless at the global level of transcripts, off-target activity can affect biological functions and phenotypes. Cell death 12 and inhibition of cell growth 10 have been reported as a direct consequence of off-target repression 13 . miRNA-like activity also yields up to B30% of false-positive hits during siRNA screening in functional genetic studies, often causing more dramatic changes in the phenotype than on-target repression 14,15 . Several approaches have been proposed to reduce off-target effects in siRNA experiments 16 , such as the use of appropriate controls 15 , low siRNA concentrations 17 , multiple siRNAs targeting the same gene individually or as a pool 18 and bioinformatics approaches 19 . Nevertheless, such indirect methods do not definitely tackle offtarget effects, for which an absolute solution is especially necessary in the clinical application 20 . Several chemically modified nucleotides such as 2 0 -O-methyl nucleotide at position 2 (2 0 -OMe) 13 and unlocked nucleic acid at position 7 (UNA) 21 have been introduced to attenuate off-target repression, although the on-target activity was also reduced 13,21,22 . However, no approach is currently available to eliminate the miRNA-like offtarget effects.
To control off-target effects, it is important to understand how the Ago-miRNA complex recognizes targets. The most well characterized features of miRNA-target interactions are seed matches, short base-pairing at least 6 nt long with seeds (within positions 2-8; ref. 4). Seed matches are informative for prediction and identification of miRNA-like off-target sites 10,11 . Such seed-mediated interactions are well supported by recent structural studies of human Ago-miRNA 23,24 and Ago-miRNA-target complex 25 , demonstrating that nucleotides in positions 2-6 are prearranged in an A-form helical structure that makes them susceptible to base pairing (Fig. 1a). Geometrical disruption between position 6 and 7 occurs in miRNA because of the insertion of an amino acid, isoleucine (I365) from human Ago2 (refs 23,24). Stepwise processes, in which nucleotides in positions 2-5 are initially exposed and paired for conformational changes to expose nucleotides 2 to 8, are required to overcome the kink in target recognition 25 .
In support of this implication, a requirement of consecutive base pairs in positions 2-6 (termed 'nucleation') was elucidated by genome-wide biochemical analysis of in vivo miRNA-target interactions 26 . This method applies cross-linking immunoprecipitation of the RNA-protein complex (CLIP) 27 coupled with high-throughput sequencing (HITS) 28 to Ago (Ago HITS-CLIP) 29 . Ago HITS-CLIP analyses that were performed in the mouse brain initially identified substantial numbers of non-canonical miRNA target sites called 'nucleation bulges', which form a bulge in target mRNAs between position 5 and 6 of the corresponding miRNA 26 . This was further identified as a general rule governing nucleation bulges, 'pivot pairing rule'. This rule determines nucleotide composition in the bulge position, postulating that nucleotide in a bulge should be able to pair with a nucleotide in position 6 (named 'pivot', Fig. 1b) 26,30 . Implicated together as the 'transitional nucleation model', nucleation bulges should transiently form consecutive base pairs up to the pivot (transitional nucleation). This is a prerequisite for initiation and propagation of base-pairing toward the 3' end for functional miRNA-target interactions, where the nucleotide originally matching the pivot in the target RNA becomes bulged-out 26,31 . Nucleation bulge sites have been observed in the mouse neocortex 29 , human brain 32 and cell lines 26,33 , also by using a ligation based Ago-CLIP method, CLASH 34 .
Consistent with the structural observation [23][24][25] , transitional nucleation may serve as a general mechanism for initiating miRNA-like target recognition 26 . This notion indicates that pairing in the pivot (position 6) serves as a decisive border to form functional miRNA-target interactions. To avoid the initiation step of miRNA-like off-target recognition, we impaired the pivot in siRNAs by substituting it with a spacer that contains no base. Such abasic pivot substitution is generated by using dSpacer (abasic deoxynucleotide; 6pi) or C3 spacer (three-carbon spacer; 6c3). Abasic pivot substitution eliminates seed-mediated off-target repression while preserving superior on-target activity (B80-100% of maximal silencing activity, relative to the siRNA without a spacer). This provides a general means for harnessing the specificity of RNAi, thus preventing potential misinterpretations of gene silencing studies or adverse effects for therapeutic applications.

miRNA-like off-target repression of siRNA is widespread.
To confirm the observation of widespread miRNA-like off-target repression 8,9 possibly mediated through transitional nucleation in siRNAs (Fig. 1b), we first performed transcriptome-wide analysis in compiled transcript profiles 35 from 35 different siRNAs (Supplementary Table 1a). In analyses of cumulative distributions of siRNA-dependent repression, the transcripts containing nucleation bulges showed a propensity for downregulation relative to the distribution of transcripts without seed matches or nucleation bulges ('No site' in Fig. 1c, left panel, and Supplementary Fig. 1a). Although nucleation bulge sites had less effect than seed sites, such downregulation was significant at sites in 3 0 -untranslated regions (3 0 -UTRs, P ¼ 3.8 Â 10 À 62 , Kolmogorov-Smirnov test (KS-test), relative to 'No site', Fig. 1c, left panel), or at any location in the transcripts (P ¼ 2.4 Â 10 À 51 , KS-test, Supplementary Fig. 1a). Analysis of a subset of the compiled data 35 focusing on siRNAs targeting a specific gene (mitogen-activated protein kinase 14; Mapk14, Supplementary Table 1b) indicated that such off-target effects mediated by nucleation bulges are concentration-, time-and sequence-dependent as evident in seed sites (Fig. 1c, right panel). These data suggest that off-target repression is only mediated via specific sequence of siRNA and not a consequence of the target gene repression.
Transcript profiles of siRNAs with 2 0 -OMe were then examined (seven different siRNAs, Supplementary Table 2) to determine whether widespread off-target repression can be effectively prevented 13 . A cumulative analysis of the putative off-target transcripts showed significant siRNA-dependent repression (Fig. 1d, Po0.05) 13 , providing evidence that siRNAs with 2 0 -OMe still caused significant off-target repression (Fig. 1d, right panel) not only for seed sites (P ¼ 2.7 Â 10 À 13 , KS-test), but also for nucleation bulge sites (P ¼ 7.4 Â 10 À 5 , KS-test). However, some off-target repression was reduced by 2 0 -OMe relative to unmodified siRNAs ( Fig. 1d and Supplementary  Fig. 1b, c). Nucleation bulges and seed sites are widespread in siRNA off-targets, and such miRNA-like off-target repression is not eliminated completely by the conventional 2 0 -OMe modification.
Abasic substitution within nucleation region of siRNAs. Widespread nucleation bulges in off-target repression ( Fig. 1) suggest that destabilizing transitional nucleation (base pairs in position 2-6) might completely block the initiation of off-target recognition. The 2 0 -OMe possesses intrinsic limitations in blocking off-target repression because it is introduced into the nucleotide backbone rather than into the base. For this reason, we used an abasic spacer, which contains no base but functions as a linker (for example, abasic deoxynucleotide spacer; dSpacer, f, Fig. 2a) as a substitute (dSpacer substitution; pi) for a nucleotide in the nucleation region (position 2-6). We hypothesized that such abasic substitution (for example, dSpacer substitution for position 6, 6pi; Fig. 2b and Supplementary Fig. 2) could effectively prevent miRNA-like off-target repression (left panel, Fig. 2b) and conserve on-target activity (right panel, Fig. 2b).
The potential for miRNA-like off-target repression was confirmed for siRNA that targets Renilla luciferase (siRL), showing significant off-target repression either through seed sites (81% repression, relative to non-targeting control (NT), Po0.01, t-test, Fig. 2c) or nucleation bulge sites (34% repression, Po0.01, t-test, Supplementary Fig. 3a) in luciferase reporter assays. Of note, siRL is often used as a negative control in siRNA experiments. A nucleotide in the nucleation region was then substituted with dSpacer (pi) to maintain the nucleotide structure without a base. Regardless of the position, all siRLs containing dSpacer substitution (pi) in the nucleation region (2pi, 3pi, 4pi, 5pi or 6pi) or position 7 (7pi) showed derepression for seed sites ( Fig. 2c and Supplementary Fig. 3a).
The dSpacer substitution (pi) may also affect the efficiency of siRNA on-target activity. Alternatively, a single mismatch generated by an abasic spacer could be tolerated by compensatory near-perfect matches (for example, dSpacer pivot substitution: 6pi; Fig. 2b, right panel). The efficiency of on-target repression was estimated by measuring half maximal inhibitory concentration (IC 50 ) through luciferase reporter assays ( Fig. 2d and Supplementary Fig. 3b,c). Every siRL containing dSpacer in the nucleation region (position 2-6) showed some repression activity (maximal inhibition rate (I max ) ¼ 43-80%, relative to the unmodified (I max [WT] ¼ 100%)). Among them dSpacer pivot substitution (6pi) yielded the best efficiency of on-target repression ( Supplementary Fig. 3b,c). Intriguingly, significant repression of seed-containing off-targets (I max ¼ 81%, IC 50 ¼ 0.78 nM, Fig. 2e), which was observed in every concentration range generally used for siRNA transfection  Transitional nucleation   3′---CAUCCUUCA ---5′   UAGG A   6  2   3′---CA U CCU CA ---5′   UAGGA GUA --3′   6  2 UAGG-A GUA --3′    Fig. 2e). By expanding the application to another siRNA (siPCSK9-A1) 36 that targets proprotein convertase subtilisin/ kexin type 9 (PCSK9), seed-mediated off-target repression (54% repression, Po0.01, t-test, Fig. 2f) was observed to be disappeared when dSpacer substitution (pi) was applied to the pivot (6pi) or other positions in the nucleation region ( Fig. 2f; IC 50 [WT] ¼ 0.01 nM versus IC 50 [6pi] ¼ ND, Supplementary  Fig. 3d). siPCSK9-A1 was previously developed to lower low-density lipoprotein cholesterol in plasma by inhibiting PCSK9-mediated degradation of the low-density lipoprotein receptor in the liver 36 . The same positional effect of dSpacer substitution on the on-target activity of siPCSK9-A1 was confirmed by observing that the dSpacer pivot substitution (6pi) outperformed others (I max [6pi] ¼ 100%, Fig. 2g and Supplementary Fig. 3e). siPCSK9-A1-6pi showed superior conservation of on-target activity (95-100% repression) without any seed-mediated off-target repression (IC 50 [6pi] ¼ ND versus IC 50 [WT] ¼ 0.01nM, Fig. 2h). This effect was especially consistent in the concentration range appropriate for therapeutic use of siRNA (0.1-1.0 nM, shaded in grey colour, Fig. 2h), although overall B2-fold decrease in the efficacy of on-target activity was observed (IC 50 [6pi] ¼ 3.4 Â 10 À 3 versus IC 50 [WT] ¼ 1.8 Â 10 À 3 nM, Fig. 2g,h and Supplementary  Fig. 3e). Of note, significant seed-mediated off-target repression was also observed at a low concentration of siPCSK9-A1 (25% repression at 0.0001 nM, Po0.01, t-test, Fig. 2h). Furthermore, dSpacer pivot substitution (6pi) did not alter slicing activity on perfectly matched on-target site (Fig. 2i), confirmed by in vitro Ago2 cleavage assay for let-7 (upper panel) and siPCSK9-A1 (lower panel). Therefore, the most favourable Perfect match site On-target Seed site Figure 2 | Effect of abasic spacer substitution for a nucleotide in the nucleation region of siRNA. (a) Abasic deoxynucleotide (dSpacer, f), applied to a nucleotide in siRNA as abasic spacer substitution. (b) dSpacer substitution (pi) in the nucleation region causes a single mismatch to seed sites in off-target transcripts (for example, dSpacer pivot substitution, 6pi), leading to unstable transitional nucleation (for example, siRL-6pi). However, siRNA-6pi may induce a stable interaction only for on-target with a perfect match site through compensatory near-perfect matches (right panel). Details are in Supplementary Fig. 2. Of note, the nomenclature 'pi' is derived from 'f' which here stands for abasic spacer substitution with a deoxynucleotide linker, dSpacer. (c) Luciferase reporter assays for miRNA-like off-target repression, mediated by seed sites for siRL (75 nM) with pi. Relative activity (average Renilla luciferase activity normalized to firefly luciferase) was analysed as a percentage relative to the control ('NT', non-targeting control siRNA); error bars, s.d. WT indicates the unmodified siRNA (red bar). Asterisk denotes Po0.01 (t-test, n ¼ 6). RL 0 indicates a different sequence of Renilla luciferase gene, that could not be targeted by siRL. (d) The same luciferase assay as in c except for measuring on-target repression (inner set). Repression efficiency was measured at different concentrations of siRL with pi (outer set; 2-6pi, indicated by different colours). IC 50 and I max values are represented in Supplementary Fig. 3c. (e) On-target activity (solid line) was examined together with off-target activity (dotted line) for siRL (red) versus siRL-6pi (blue) by luciferase reporter assays. General siRNA concentrations used for cell culture are indicated (grey, 1-100 nM). (f-g) Effects of dSpacer substitution (pi) in the nucleation region were also examined for siPCSK9-A1 (ref. 36) in f as in c and in g as in d. (h) The same analysis as in e except for siPCSK9-A1; the grey colour indicates the therapeutic concentration. (i) In vitro Ago2 cleavage assays for let-7 (upper panel) and siPCSK9-A1 (lower panel). The triangle denotes expected size of cleaved product from the target substrate (indicated with a line). position for abasic substitution in siRNAs is the pivot (position 6) because this abolishes seed-mediated off-target repression while maintaining on-target activity.
dSpacer pivot substitution outperforms in target specificity. We also investigated the possibility that other conformations of the abasic spacer could outperform the dSpacer pivot substitution (6pi). First, the effect of rSpacer (abasic ribonucleotide) substitution (pi-r) in the nucleation region of siRL was examined (Fig. 3a,b and Supplementary Fig. 4). Abolition of the seedmediated off-target repression was observed when the rSpacer substitution was applied to a nucleotide in positions 3-6 ( Fig. 3a). Nevertheless, all such siRLs containing the rSpacer substitution had less efficient on-target activity (I max r70%, IC 50 Z0.99 nM) than dSpacer pivot substitution (6pi, I max ¼ 80%, IC 50 ¼ 0.36 nM; Fig. 3b and Supplementary Fig. 4). Next, the insertion of dSpacer (pi-b) into the nucleation region was noted to cause a bulge with alignment to the target site (Fig. 3c, upper panel). The seedmediated off-targets were found to be derepressed when dSpacer insertion was within positions 4-6 ( Fig. 3c, lower panel). Although some on-target activity was retained (I max r48%, IC 50 Z3.06 nM), all of these insertions showed less efficient ontarget activity than dSpacer pivot substitution ( Fig. 3d and Supplementary Fig. 5). Similarly, after insertion of rSpacer (pi-rb), abrogation of off-target repression was observed (positions 4-6, Fig. 3e) with insufficient preservation of on-target activity (I max r55%, IC 50 Z0.81 nM) in comparison with the dSpacer pivot substitution ( Fig. 3f and Supplementary Fig. 6). All of these data demonstrate that although off-target repression can be abolished by other spacer conformations, which used rSpacer or caused an abasic bulge of siRNA in siRNA:target duplex, these conformations cannot maintain on-target activity as well as dSpacer pivot substitution.
Abasic pivot substitution eliminates off-target repression. To further confirm whether abasic pivot substitution can abolish miRNA-like off-target recognition, we assessed seed-mediated off-target activity for nucleation bulge sites of PCSK9-A1 and siRL (Fig. 4a,b). Significant off-target repression mediated by nucleation bulge sites for PCSK9-A1 (IC 50 ¼ 0.08 nM, 63% repression, Po0.01, t-test, Fig. 4a) and siRL (IC 50 ¼ 0.61 nM, 38% repression, Po0.01, t-test, Fig. 4b) was abolished by dSpacer pivot substitution (6pi, IC 50 ¼ ND, 0% repression, Fig. 4a,b). In addition, siRL-6pi showed inhibition of passenger strandmediated off-target activity ( Supplementary Fig. 7). The use of dSpacer pivot substitution was then tested on miRNAs with the expectation of abrogating seed-mediated target repression (Fig. 4c,d). Synthesized miRNA duplexes containing the dSpacer pivot substitution showed derepression of the target with seed sites (IC 50 ¼ ND, 0% repression) for miR-708 (IC 50 ¼ 0.05 nM, 20% repression, Fig. 4c) and cel-miR-67 (IC 50 ¼ 0.31 nM, 23% repression, Fig. 4d), the C.elegans-specific miRNA often used as a control. Notably, in all ranges of concentration tested here, there was no significant change in relative activity of luciferase reporters for every siRNA and miRNA with dSpacer pivot substitution (IC 50 [6pi] ¼ ND, 0% repression, Figs 3 and 4). All of these data provide evidence that the dSpacer pivot substitution is generally applicable to any siRNA for eliminating miRNA-like off-target repression. Abasic pivot substitution conserves on-target activity. To examine the extent of preserved on-target activity in siRNA-6pi, IC 50 and I max were also measured for a perfectly matched site of miR-124 (Fig. 5a). Luciferase reporter assays showed that miR-124-6pi has the same maximal repression activity as miR-124 (I max ¼ 100%), although application of dSpacer pivot substitution showed slight reduction (B2-fold decrease, Fig. 5a). For the siRNA targeting MAPK14 (siMAPK14), near-perfect conservation of the on-target activity was observed (I max ¼ 100%) across all ranges of concentrations used for cell cultures (100% conservation, 0.05-75 nM, Fig. 5b). This was further confirmed by immunoblot analyses (Fig. 5c and Supplementary Fig. 8a,b). When the dSpacer pivot substitution was applied to siPCSK9-A2, which has the same sequence as siPCSK9-A1 but with 2 0 -OMe modification to increase stability and avoid innate immune responses in vivo 36 , the same conservation of on-target activity was observed (I max ¼ 100%, Fig. 5d, left panel). Indeed, both siPCSK9-A2 and siPCSK9-A2-6pi efficiently silenced PCSK9 mRNA by inducing the same degree of repression (B5-fold, Fig. 5d, right panel). This finding was further confirmed for siPCSK9-A1 by immunoblotting analysis ( Supplementary  Fig. 8c). Importantly, all of these small RNAs showed abolition of miRNA-like activity when they contained the dSpacer pivot substitution (IC 50 ¼ ND, 0% repression, Fig. 5e-h). Taken together, these results lead to the conclusion that dSpacer pivot substitution maintains superior on-target activity (B80-100%, relative to the unmodified) while avoiding the off-target repression.
Incomplete elimination of off-targets by conventional methods. Next, we examined the efficiency of blocking the miRNA-like off-target repression by conventional methods such as 2 0 -OMe 13 and UNA 21 . For this purpose, we applied 2 0 -OMe or UNA to miR-124, which has the most Ago-bound seed sites in the brain 29,32 , to see whether its inhibitory effect is strong enough to block the silencing activity of this seed-centric miRNA (seed, IC 50 ¼ 0.07 nM, I max ¼ 52%, relative to perfectly matched target, Fig. 5e; nucleation bulge, IC 50 ¼ 0.78 nM, I max ¼ 27%, Supplementary Fig. 8d). In the luciferase reporter assay, miR-124 containing 2 0 -OMe showed significant repression for both seed (IC 50 ¼ 0.65 nM, I max ¼ 32%, Fig. 5e and Supplementary Fig. 8d) and nucleation bulge sites (IC 50 ¼ 0.91 nM, I max ¼ 19%, Supplementary Fig. 8e). Such incomplete inhibition of seedmediated miRNA-like activity was also observed for UNA applied to miR-124 (IC 50 ¼ 7.2 nM, I max ¼ 21%, Fig. 5e) and siPCSK9-A2 Fig. 5h). Bulge-siRNA, which was developed to alleviate off-target repression by containing a bulge at position 2 of the guide strand in the siRNA duplex 37 , also showed remaining off-target activity when applied to siPCSK9-A2 (IC 50 ¼ 0.96 nM, I max ¼ 29%, Fig. 5h). However, the dSpacer pivot substitution showed elimination of seed-mediated off-target repression in all cases (IC 50 ¼ ND, 0% repression, Fig. 5e-h). This finding was further confirmed by immunoblotting of PTBP1, a previously validated miR-124 target with seed sites 38 ( Fig. 5f and Supplementary Fig. 8f- Supplementary Fig. 9a). In addition, the dSpacer pivot substitution slightly outperformed UNA in on-target repression by siPCSK9-A2 ( Supplementary Fig. 9b).
Overall, the current methods for overcoming off-target silencing have limited potency, whereas dSpacer pivot substitution could eliminate miRNA-like off-target repression with improved potency for preserving on-target repression.
Most of the miR-124 targets are known to function in neuronal differentiation 29 , explaining the neurite outgrowth phenotype induced by expression of miR-124 (ref. 39) in neuroblast-like cells such as Neuro2a (N2a) 38,40 . As disturbances of miR-124 function were expected in miR-124-6pi, this possibility was first confirmed by integrated analysis of de novo Ago-miR-124 clusters 26,29 and RNA-Seq data on biological pathways, with focus on regulation of the actin cytoskeleton ( Supplementary Fig. 11a). One of the wellvalidated critical miR-124 targets, integrin b-1 (ITGB1) 29,39 , was analysed (Fig. 6d). Consistent with the results (Figs 5e,f), ITGB1 mRNA was derepressed completely by the dSpacer pivot substitution but only marginally by 2 0 -OMe (Fig. 6d). In addition, RNA-Seq analysis demonstrated that 2 0 -OMe could not sufficiently block the global repression of off-targets as much as the dSpacer pivot substitution in siRL (Fig. 6e). In N2a cells, miR-124-6pi was found to lose its function of inducing neurite outgrowth (Fig. 6f). In contrast, miR-124 containing 2 0 -OMe still induced neurosphere-like structures and differentiation (Fig. 6f, Supplementary Fig. 11b and Supplementary Movies 1-3). These data showed that abasic pivot substitution abolishes miRNA-like target interactions and repression, enough to prevent miRNA-induced biological phenotypes.
To explore the expected phenotypic consequences of this off-target repression, gene ontology (GO) analysis was performed and revealed that functional categories related to 'metal ion  Table 4). Some of these target genes are known to regulate copper metabolism (metallothionein 1 and metallothionein 2; Supplementary Table 4). Consistent with these findings, the concentration of intracellular copper (17±4.1 mg dl À 1 ) was significantly increased depending on siPCSK9-A2 expression (25 ± 2.2 mg dl À 1 ) in the mouse liver cell line (NCTC clone 1469, Fig. 7e). These effects are similar to those caused by exposure to increased plasma copper concentration (B32 mM) in Wilson disease 41 (22±3.1 mg dl À 1 , Fig. 7e). siPCSK9-A2 also significantly induced apoptotic cell death (B19% increase, Po0.01, t-test, Fig. 7f and Supplementary  Fig. 12d) as observed in copper-induced apoptosis ( Supplementary Fig. 12c) 41 . The same off-target phenotypes were also observed with siPCSK9-A1 ( Supplementary Fig. 12b,c). Moreover, all of these off-target phenotypes disappeared when they contained the dSpacer pivot substitution (Fig. 7e,f and Supplementary Fig. 12b-d).
As miRNA-like off-target effects were reported to be species-specific 42 , the human liver cell line (HepG2) was used to evaluate the potentially deleterious off-target effects and their possible prevention by the abasic pivot substitution in humans, a clinically important application. RNA-Seq analysis for transcript profiles depending on siPCSK-A2 expression (Supplementary  Table 3c) showed widespread off-target transcripts in human liver cells (Fig. 7g), whereas the dSpacer pivot substitution significantly derepressed such off-targets (P ¼ 1.1 Â 10 À 9 , KS-test, Supplementary Fig. 13a). This included putative direct miRNA-like off-targets that contain seed sites (P ¼ 0.01, KS-test, Supplementary Fig. 13b).
To determine the off-target phenotypes, the GO functions of the derepressed transcripts were also analysed ( Supplementary  Fig. 13c,d). Cell cycle regulation was identified as the most likely affected phenotype, based on comparisons of fold changes between siPCSK9-A2 and siPCSK9-A2-6pi ( Fig. 7h and Supplementary Fig. 13e,f). This finding was confirmed by functional pathway analysis (Supplementary Fig. 14a). Cell cycle analysis of siPCSK9-A2-expressed HepG2 showed that B12% of the cells were significantly decreased in G1/S but increased in G2/M, whereas siPCSK9-A2-6pi did not cause these phenotypes (Fig. 7i). The same results were observed with siPCSK9-A1 ( Supplementary Fig. 14b) and such off-target effects were found to be species-specific (Supplementary Fig. 15). Taken together, we conclude that unexpected off-target phenotypes are inevitably caused by siRNAs in vivo, but the dSpacer pivot substitution can also be used in vivo to eliminate such adverse off-target effects.

Discussion
RNAi has now become one of the most widely used methods for gene silencing because of specific inhibition of gene expression and ease of use 1 . It has been applied to study various genes for loss of function and develop therapeutic applications to knock down disease-causing genes 20 . Our initial assumption about the specificity of RNAi was found to be relative, only indicating that on-target repression activity is stronger than off-target repression 8 . Chemical modifications such as 2 0 -OMe 13 and UNA 21 have been utilized to reduce the off-target effects, but their inhibitory effect on miRNA-like off-targets is limited, as we found in this study. This may be because such modifications have been applied only to the nucleotide backbone rather than to the bases, which directly participate in target recognition. On the basis of this notion, the effectiveness of abasic substitution in blocking off-target activity and selecting specific alleles 43 can be explained.
As there is a trade-off between on-target activity and derepression of off-targets, a critical issue regarding any method that blocks off-targets is how effectively on-target activity can be maintained. Structural studies showed that the base and 2 0 -hydroxyl group (2 0 -OH) of a pivot nucleotide are located at a close distances from the a-helix of human Ago (base, 3.6 Å from I365; 2 0 -OH, 3.9 Å from A369), which generates a kink after the pivot position (positions 6-7, Fig. 1a) 23,24 . Analysis of models derived from the human Ago-miRNA structure 23 elucidated that dSpacer, which contains neither a base (4.9 Å from I365) nor 2 0 -OH (5.1 Å from A369), has less potential to cause steric hindrance ( Fig. 8a and Supplementary Fig. 16a-c). In a ternary structure of the human Ago2-RNA-target, the kink in Ago-miRNA structure is absent by moving out the a-helix and a perfectly matched target widens central cleft of Ago to accommodate target pairing beyond position 8 (ref. 25). On the basis of the model derived from the ternary structure 25 , dSpacer pivot substitution seems to provide adequate space to be beneficial for such stepwise conformational transitions: moving out the kink and widening the central cleft (Fig. 8b), requiring to be further stabilized by pairing to position 8 and/or position 13-16 25 . Although 2 0 -OH of pivot nucleotide does not make a hydrogen bond interaction [23][24][25] , positions 2 to 7 (including pivot) of the miRNA-target duplex make extensive hydrophobic and van der Waals interactions with aliphatic segments within a-helix and PIWI domain 25 . Therefore, removing 2 0 -OH of pivot nucleotide may increase Ago target affinity, resulting in conservation of on-target activity. In support, other abasic conformations, which either involve rSpacer (6pi-r, Fig. 3b) or cause an abasic bulge (6pi-b, Fig. 3d; 6pi-rb, Fig. 3f), showed less efficient on-target activity than dSpacer pivot substitution.
This notion was further examined by using C3 spacer (Fig. 8c, upper panel) which consists of three carbons minimally required for the linker function (Fig. 8c), found to have the least bulky structure without 2 0 -OH in pivot (6.4 Å from I365, 6.2 Å from A369, Supplementary Fig. 16d). As expected, C3 spacer substitution elicited excellent repression activity for a perfectly matched on-target ( Fig. 8c and Supplementary Fig. 17a,b) without inducing seed-mediated target repression ( Fig. 8d and Supplementary Fig. 17c,d). Therefore, destabilizing nucleation pairings by the abasic pivot substitution may enable siRNA to avoid miRNA-like target recognition (Fig. 8e, upper panel), but facilitate to transit from weak nucleation to functional interaction with the perfectly matched on-target (Fig. 8e, lower panel), especially in the case where the abasic pivot has reduced potential of steric hindrance via abasic and no 2 0 -OH (for example, dSpacer, C3 spacer).
Although abasic pivot substitution conserves the on-target activity of siRNAs (B80-100% of maximal silencing activity, relative to siRNAs without a spacer), there is some variation in its efficiency among different siRNAs (r9-fold decrease in IC 50 , Figs 2 and 5). Such variation of on-target activity possibly depends on the sequence composition and secondary structure of target mRNAs, which is reminiscence of the general  Table 4). (e) The amount of intracellular copper in NCTC clone 1469 cells was significantly increased by siPCSK9-A2 (25 ± 2.2 mg dl À 1 , Po0.01, relative to NT, t-test, n ¼ 3), but not by siPCSK9-6pi (16±5.0 mg dl À 1 ). 'Cu 2 þ ' indicates treatment with 32 mM CuSO 4 . A single asterisk denotes Po0.05 (t-test). (f) Cell death assays of NCTC clone 1469 cells measured by FACS analysis with propidium iodide (PI) and Annexin V staining. The percentage of cells in the phase of early apoptosis (red box, mean ± s.d., n ¼ 3) was significantly increased by siPCSK9-A2 but not by siPCSK9-A2-6pi (details in Supplementary Fig. 12d).
(g) The same transcriptome-wide analysis of off-targets as in d except on HepG2 cells. (h) Comparison of fold changes between siPCSK9-A2 and siPCSK9-A2-6pi, in a set of off-target transcripts functioning in cell cycle regulation, identified by GO analysis in 6pi-dependent derepressed transcripts (dotted box in g, Supplementary Fig. 13c-f). (i) Cell cycle analysis of HepG2 cells, performed by FACS analysis using PI staining. A defect in cell cycle regulation was observed with siPCSK9-A2, but not with siPCSK9-A2-6pi (n ¼ 3).
properties of RNAi 1 . Most of siRNAs containing abasic pivot substitution showed the same on-target activity (100%) as unmodified versions in the concentration range generally used for siRNA transfection into cell cultures (10-75 nM, Figs 2 and 5; except for siRL of which on-target activity is B80%). However, further studies should be performed to determine features of siRNA sequences, which perform the same on-target activity even when they contain abasic pivots (Fig. 5b,d), an important issue for clinical application to avoid dose-limiting toxicity. By substituting pivot with an abasic spacer, no significant miRNA-like off-target repression was observed in every concentration range or even the maximal concentration of siRNA transfection (100 or 150 nM, Supplementary Fig. 18a). dSpacer pivot substitution did not change the amount of siRNAs loading into Ago (Supplementary Fig. 18b,c), retained the ability to cleave (Fig. 2i and Supplementary Fig. 18f) and degrade on-target mRNAs ( Supplementary Fig. 18d,e). Although miRNA-like offtarget effects were eliminated by the abasic pivot substitution, there was still remaining off-target effect causing an innate immune response against double-strand RNAs-the dSpacer pivot substitution was observed to have no effect on TLR3-mediated innate immune responses ( Supplementary  Fig. 18g). Therefore, it is suggested that abasic pivot substitution should be combined with some other modification methods that can prevent such innate immune responses, as we tried and observed in the case of siPCSK9-A2-6pi (Figs 5d,h and 7).
The abasic pivot substitution is broadly applicable to a wide range of RNAi usages. Although all of the siRNA sequences used in this study were designed with consideration of off-targeting 16,19,44 , they all showed significant off-target repression. With long-term therapeutic application, the off-target effects of siRNA may become serious, as implicated in this study of PCSK9 siRNAs. The abasic pivot substitution is essential to apply RNAi  Fig. 1a (from 4F3T 24 in PDB). Of note, 6pi reduced steric hindrance by generating space (4.9 Å from I365, 5.1 Å from A369) in the kink between position 6 and 7 (3.6 Å from I365, 3.9 Å from A369, Fig. 1a). Details are provided in Supplementary Fig. 16. (b) A surface model of Ago-miRNA-target structure with 6pi (from 4W5O 25 in PDB). Target mRNA is yellow. (c) The C3 spacer (upper panel) substitution for pivot (6c3) was applied to miR-124 and its effect on repressing perfectly matched sites was analysed by estimating IC 50 using luciferase reporter assays, as in Fig. 5a. (d) Effect of 6c3 on seed-mediated target repression was measured for miR-124 by luciferase reporter assays (IC 50 [6c3] ¼ ND, IC 50 [WT] ¼ 0.07 nM) as in Fig. 5e. ND, not determined. Asterisk denotes Po0.01 (t-test, n ¼ 6). (e) miRNA-like off-target sites mediated by siRNA seed regions may be unable to induce stable transitional nucleation owing to abasic pivot substitution, avoiding miRNA-like off-target recognition and repression (upper panel). In contrast, the on-target site that is perfectly complementary to siRNA except for the abasic pivot overcomes unstable transitional nucleation (only four consecutive base pairs in positions 2-5) and the structural kink by 3 0 compensatory pairing with help of the reduced steric hindrance (as modelled in b). This event leads to formation of the siRNA-target duplex and on-target repression (lower panel). Shaded ovals represent the Ago protein.
for experimental and clinical purposes, where ensuring the specificity is especially important.
Prediction of miRNA-like target sites. Basically, all 6mer matches to seed sequences in position 1-8 were searched to identify seed sites (6-8 nucleotides) in 3 0 -UTR or whole transcripts (defined by RefSeq, downloaded from the UCSC genome browser). To search nucleation bulge sites, 7mer matches were derived from 6mer seed matches (position 2-7) in which the nucleotide of the position 5-6 target mRNA bulge sequence is complementary (Watson-Crick base pairing) to position 6 (pivot) of the corresponding siRNAs or miRNAs, as described previously 26,30 . For comparison of the seed and the nucleation bulge sites, we used only 7mer seed matches with lengths that are the same as those of 7mer nucleation bulge patterns.
Meta-analysis of microarray data. Meta-analysis was performed by obtaining normalized compiled data 35 13 . For meta-analysis regarding 2 0 -OMe, normalized microarray data from the expression of seven different sequences of siRNAs as pairs (no modification versus 2 0 -OMe) were downloaded from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/), in entirely or selectively for transcripts with significant differences (Po0.05). To select potent siRNA off-target transcripts, we predicted miRNA-like off-target sites (seed sites or nucleation bulge sites). Cumulative fraction analyses depending on fold change (log 2 ratio) were performed only for coding transcripts (RefSeq IDs start with NM) and were analysed as described previously 26,29 . KS-tests were performed by using Scipy (scipy.stats.ks_2samp()).
Cell culture and transfection. The human cervical adenocarcinoma cell line HeLa (ATCC CCL-2), human hepatocellular carcinoma cell line HepG2 (Korean Cell Line Bank) and mouse neuroblastoma cell line Neuro-2a (N2a, ATCC CCL-131) were maintained in Dulbecco's modified Eagle's medium (DMEM; Gibco) supplemented with 10% fetal bovine serum (FBS, Gibco), 100 U ml À 1 penicillin, and 100 mg ml À 1 streptomycin at 37°C with 5% CO 2 incubation. The mouse liver cell line NCTC clone 1469 (Korean Cell Line Bank) was maintained either in the same medium or with 10% horse serum (WelGENE) instead of FBS. NCTC clone 1469 cells were transfected by using Lipofectamine 2000 (Invitrogen), whereas HepG2 and N2a cells were transfected by using RNAiMAX (Invitrogen) with 50 nM RNA duplexes according to the general protocol provided by the manufacturer, unless otherwise indicated. Transfection into HeLa cells was performed as described previously 26 . The cells were generally collected 24 h after transfection in all experiments, unless otherwise indicated.
Luciferase reporter assays. Luciferase reporter assays were performed as described previously 26 . In brief, psiCheck-2 plasmids (Promega) were cotransfected with duplexed siRNAs or miRNAs by using Lipofectamine 2000 (Invitrogen). Twenty-four hours after transfection into HeLa cells, relative activity (Renilla luciferase activity normalized to firefly luciferase) was measured by Dual-Luciferase Reporter Assay System (Promega) with the GloMax-Multi Detection System (Promega) with replicates (n ¼ 6) according to the manufacturer's protocol. In general, IC 50 was calculated by performing nonlinear least squares fitting for the sigmoid function using Scipy (scipy.optimize.curve_fit()). In cases where least squares failed to fit the function, an approximate IC 50 was calculated from the regression line.
Northern blot for small RNAs. Total RNA was extracted using TRIzol (Invitrogen) and digested with DNase 1 (Qiagen). 10 mg RNA was incubated with formamide (Sigma-Aldrich) at 65°C for 10 min and separated in 12% acrylamide denaturing (urea) gels. The gels were stained with SYBR-Safe (Invitrogen) to visualize the RNA samples. RNA was transferred to Zeta-Probe GT membranes (Bio-Rad) and cross-linked by ultraviolet irradiation (Stratagene). The membranes were blocked in hybridization buffer (0.5 M Na 2 PO 4 pH7.2, 15% formamide, 1% BSA and 7% SDS) for 30 min at 37°C in a rotating hybridization oven. The miR-124 probes (5 0 -GGCATTCACCGCGTGCCTTA-3 0 ) were labelled with g-32 P ATP using PNK (NEB) and cleaned with a G25 column (GE Healthcare). Probes were heated at 65°C before being added into the hybridization bottle, and membranes were hybridized at 37°C overnight. The membranes were washed three times with 1 Â SSC containing 0.1% SDS, and then was exposed to X-ray film for 1 day at À 80°C.
Cell cycle and cell death analyses. For cell cycle analysis, HepG2 or NCTC clone 1469 were harvested at 48 h after transfection, resuspended (1 Â 10 5 cells) in 200 ml of PBS containing 20 mM EDTA (Sigma-Aldrich), and then fixed with 700 ml of ethanol (Biosesang) for an hour at 4°C. Then, 10 ml of 1 mg ml À 1 propidium iodide (PI; Sigma-Aldrich) was added after treatment with 5 ml RNase A (0.2 mg ml À 1 , Sigma-Aldrich) for 30 min at 37°C. The cells were analysed by BD FACSCalibur (BD Biosciences). Of note, all the cells were synchronized by 24 h serum starvation after the transfection.
For cell death analysis, FITC Annexin V Apoptosis Detection Kit I (BD Pharmingen) was used. In brief, B1 Â 10 5 cells in 100 ml of 1 Â binding buffer were stained by adding 5 ml of 50 mg ml À 1 PI and 5 ml of FITC Annexin V for 15 min. After adding 400 ml of 1 Â binding buffer, the samples were transferred to 5 ml tubes with cell strainer caps (BD Falcon) and analysed by BD Aria I (BD Biosciences). In this assay, the cells were collected 72 h after the transfection.
Experiments with mice. siRNAs (5 mg kg À 1 ) were administered to 6-week-old male C57BL/6 mice (Orient, a branch of Charles River Laboratories) via intravenous injections using in vivo-jetPEI (Polyplus) according to the manufacturer's protocol. The mice were divided into three groups depending on the injected siRNAs (n ¼ 5 each group, sample size was chosen on the basis of the minimum number used in previous study 36 ): NT, A2 and A2-6pi. Briefly, siRNA and in vivo-jetPEI complexes (N/P ratio ¼ 6) were generated by following the manufacturer's protocol and injected into the tail vein with a sterile syringe (1.0 ml) and a 30-gauge needle. Two days after the injections, liver tissues and blood from the abdominal aorta were collected. Plasma was separated from the blood by centrifugation (2,000g for 20 min at 4°C). All the samples were stored at À 80°C until analysed. Total cholesterol in murine plasma was measured by the Cholesterol E Enzymatic Colorimetric Method (Wako Chemicals) according to the manufacturer's protocol. The excised liver tissues were used for RNA extraction, followed by RNA-Seq and qPCR analysis. This study was reviewed and approved by the Institutional Animal Care and Use Committee (IACUC) of Samsung Biomedical Research Institute (SBRI). SBRI is an Association for Assessment and Accreditation of Laboratory Animal Care International (AAALAC International) accredited facility and abides by the Institute of Laboratory Animal Resources (ILAR) guide.
Quantification of intracellular copper. NCTC clone 1469 cells were washed once with D-PBS (PBS without Ca 2 þ and Mg 2 þ ; WelGENE), detached by scraper and resuspended in D-PBS. The same number of cells (B4 Â 10 4 ) was prepared by ADAM Automated Cell Counter (NanoEnTek) following the manufacturer's protocol. The cells were lysed by sonication with Bioruptor (Diagenode): 30 s on/off interval for five treatments (5 min total) at high-power settings. The cell lysates were cleared by centrifugation (12,000g for 5 min at 4°C), lysed in 0.5 N NaOH (incubation for 30 min with shaking at 1,000 r.p.m.), and boiled at 100°C for 30 min. The copper concentration of prepared samples was measured by using QuantiChrom Copper Assay Kits (BioAssay Systems) following the manufacturer's instruction. The cells were collected at 48 h after transfection. As a positive control for this assay, the cells were treated with 32 mM CuSO 4 and collected 4 h after transfection.