Small RNAs with Imperfect Match to Endogenous mRNA Repress Translation

A 21-base pair RNA duplex that perfectly matches an endogenous target mRNA selectively degrades the mRNA and suppresses gene expression in mammalian tissue culture cells. A single base mismatch with the target is believed to protect the mRNA from degradation, making this type of interference highly specific to the targeted gene. A short RNA with mismatches to a target sequence present in multiple copies in the 3′-untranslated region of an exogenously expressed gene can, however, silence it by translational repression. Here we report that a mismatched RNA, targeted to a single site in the coding sequence of an endogenous gene, can efficiently silence gene expression by repressing translation. The antisense strand of such a mismatched RNA requires a 5′-phosphate but not a 3′-hydroxyl group. G·U wobble base pairing is tolerated as a match for both RNA degradation and translation repression. Together, these findings suggest that a small inhibitory RNA duplex can suppress expression of off-target cellular proteins by RNA degradation or translation repression. Proper design of experimental small inhibitory RNAs or a search for targets of endogenous micro-RNAs must therefore take into account that these short RNAs can affect expression of cellular genes with as many as 3–4 base mismatches and additional G·U mismatches.

Small inhibitory RNAs (siRNAs) 1 are small RNAs of ϳ21nucleotide length that trigger the destruction of target mRNA with which they share complementarity (1). In vivo, they are generated by an RNase III enzyme, Dicer, which cleaves a precursor double-stranded RNA (dsRNA) into ϳ21-nucleotide sense, antisense, or, potentially, dsRNA fragments (2)(3)(4)(5). The ϳ21-nucleotide siRNA is incorporated into an RNA-induced silencing complex (RISC) that mediates the cleavage of target mRNA. RNA-mediated silencing was first observed in petunia flowers as post-transcriptional gene silencing (PTGS). Jorgensen et al. observed silencing of the introduced transgene and the endogenous gene homologous to the transgene (6,7). The authors defined this activity of exogenous elements to alter the expression of the endogenous gene as "cosuppression." The key breakthrough came with the work of Mello and co-workers, who were employing antisense RNA for gene silencing in Caenorhabditis elegans (1). They observed that the mixture of sense and antisense RNA was 10-fold more effective than single strand sense or antisense RNAs. The authors named this double-stranded RNA-mediated silencing "RNA interference (RNAi)." After their report, RNAi-mediated silencing was linked to a previously observed cosuppression phenomenon in plants, fungi, and Drosophila (6 -9). Tuschl and colleagues used RNAi to silence genes in mammalian tissue culture cells (10). This was significant, because the authors could demonstrate sequence-specific silencing of target genes in mammalian cells without any accompanying nonspecific interferon-mediated response. Subsequently, siRNA has been used extensively to silence specific genes for experimental purposes and to target infectious agents like the human immunodeficiency virus and the hepatitis B and C viruses (11)(12)(13)(14)(15)(16).
Distinct from siRNAs are another group of small RNAs called small temporal RNAs (stRNAs), named as such because they are temporally regulated during the development of C. elegans. Also called microRNAs, they silence genes by repressing translation while not affecting the levels of mRNA (17)(18)(19)(20)(21). Important differences between siRNA and stRNA mechanisms have been noted. Double-stranded RNAs with long stretches of perfectly base-paired nucleotides are cleaved to produce smaller double-stranded siRNAs. On the other hand, double-stranded RNA with imperfect complementarity between the strands is cleaved to produce single-stranded stRNA (17). The most significant difference lies in the mechanism of gene silencing. siRNA causes degradation of a target mRNA to which it is perfectly base-paired, whereas stRNA inhibits translation of a target mRNA to which it is imperfectly base paired.
The best studied stRNAs, lin-4 and let-7, repress their target genes, lin-14, lin-28 and lin-41, by partial base pairing to multiple target sequences in the 3Ј-untranslated region (UTR) (22)(23)(24). Mutations in lin-4 and let-7 lead to developmental defects due to loss of repression of these target genes at specific stages of C. elegans development (24). Recently, a gene (bantam) has been identified in Drosophila that encodes a micro-RNA (miRNA) whose expression is temporally and spatially regulated (21). Around 100 miRNAs have now been identified in humans (18,25,26). Some of the miRNAs (let-7) and their targets (lin-41) have been reported across eukaryotes, suggesting that these small RNAs may participate in the genetic regulation of a wide range of cellular pathways (27). Because the targets of miRNA share only partial complementarity, very little is known about the targets or functions of most miRNAs.
Transfection of siRNA complementary to an mRNA is a fast and convenient tool for studying loss of function of a target gene in mammalian cells. However, there are conflicting reports in literature about the specificity of the sequence match necessary to achieve gene silencing (28 -30). Recent reports by at least two different groups have claimed high specificity of RNAi silencing by analyzing genome-wide expression profiling (28,30). However, others have observed that partial complementarity between siRNA and mRNA leads to nonspecific changes in mRNA levels, raising an interesting debate on specificity of siRNA silencing (29). Although some report that a single base mismatch with respect to a target abolishes silencing, others have observed that mismatches can be tolerated (13,(31)(32)(33)(34). Semizarov et al. have observed that scrambled siRNA (the complementary siRNA with an inversion of a four-nucleotide fragment in the middle of sequence) does not cause any change in the level of target mRNA and protein (28). Here, we have investigated the number of mutations that can be tolerated for a short RNA to suppress expression of an endogenous gene. We observed that when 3-4 mismatches are introduced into p21 or geminin siRNA, they can still silence the targeted gene, but by triggering translational repression instead of mRNA degradation. The 5Ј-phosphate group is required at the antisense strand for translational repression in human cells. The 3Јhydroxyl group is dispensable, eliminating any role of a primer extension-mediated amplification step in dsRNA-mediated translational repression silencing. The G⅐U wobble base pair is recognized like an authentic Watson-Crick base pair in antisense RNA oligonucleotide-mRNA duplex interaction. Base pairing at the center of duplex is more critical, stabilizing the antisense RNA oligonucleotide-mRNA duplex and enhancing the silencing.
A recent report showed that mismatched short RNAs can translationally silence a target mRNA with multiple targets in the 3Ј-UTR (35). Because the target mRNA was an artificial construct expressed from an exogenous promoter, it was not clear how these findings could apply to the conventional siRNA experiments in which the short RNA duplex is targeted to a single site in the coding region of an endogenous gene. We report here that a short RNA duplex similar to those used in siRNA experiments can function like an stRNA and silence an endogenous gene that carries only one natural target sequence partially complimentary to it. We also report that the target for this stRNA-like process is not limited to 3Ј-UTR and can include the coding sequence of a gene. Taken together, our results emphasize the possibility of off-target silencing by siRNAs used for an experimental or a therapeutic purpose.

EXPERIMENTAL PROCEDURES
Cell Culture, Transfections, Protein, and RNA Analysis-PC3 cells from the American Type Culture Collection (Manassas, VA) were grown in RPMI 1640 medium supplemented with 10% fetal bovine serum and antibiotics, whereas HeLa cells were maintained in Dulbecco's modified Eagle's medium. For dsRNA transfection, commercial RNA duplexes composed of 21-nucleotide sense and 21-nucleotide antisense strands were used, paired so as to have a two-nucleotide 3Ј-overhang (Dharmacon and Ambion). Cells were grown to ϳ50% confluence in a 6-well plate and transfected three times at 24-h intervals with 0.32 nmol of annealed RNA duplex (4.2 g of RNA duplex) using Oligofectamine reagent (Invitrogen). Transfection was done in the absence of fetal bovine serum, which was replenished 4 h after each transfection. By using this protocol and Oligofectamine reagent, cell mortality was Ͻ10%, even after three dsRNA transfections. The cells were then assayed for levels of targeted protein and mRNA. A nonspecific RNA duplex (GL2), complimentary to firefly luciferase (5Ј-AACGUACGCGGAAUACUUCGA-3Ј), was used as a negative control.
Cell extracts were prepared in ice-cold extraction buffer (50 mM Tris, pH 7.4, 150 mM NaCl, 0.2% Nonidet P-40, and 1 unit/ml of DNase) containing 0.1 mM phenylmethylsulfonyl fluoride and 1 mg/ml each pepstatin A, leupeptin, and aprotinin. The concentration of total protein in the lysates was determined by Bradford assay, and equal amounts were used for Western blots. The geminin antibody has been described earlier (36), whereas p21 antibody (sc-397) was obtained from Santa Cruz Biotechnology. RNA was isolated with Trizol reagent (Invitrogen). For Northern analysis of geminin and p21 mRNA, 10 g of total cellular RNA was separated on denaturing agarose gel, transferred to nylon membrane, and probed with random primed geminin and p21 32 Plabeled cDNA probes. The quantitation of mRNA levels was done with the Fuji BAS-2500 phosphorimaging device and Image Gauge software. Protein levels were estimated with NIH software-Scion Image by scanning autoradiograms.

Imperfect Base Pairing with the Target Causes siRNA to
Silence the Gene in an stRNA-like Process-We initially designed an siRNA (p21-2) against a target in the 3Ј-UTR of the endogenous p21 mRNA (Fig. 1A). Upon transfection of PC3 Consistent with the silencing being mediated by RNAi, mRNA levels were significantly reduced (Fig. 1, B and C). Next, we incorporated mismatches in the siRNA to observe the effect of the mismatch on silencing. With 3-4 mismatches (p21-4, p21-10, and p21-11) we still observed a significant decrease in p21 protein levels. However, the mRNA was not degraded (Fig. 1, A-C). These results contrast to those observed by others, wherein a scrambled siRNA (the complementary siRNA with an inversion of a four-nucleotide fragment in the middle of sequence) does not cause any changes in levels of target protein (28). Therefore a short RNA duplex with mismatches to the target repressed protein levels without degrading the target mRNA. Given the mechanism of gene silencing by stRNA, the most likely mechanism of this repression of protein levels is through translational repression.
stRNA-like Silencing Can Be Observed Even When Targeted to the Coding Sequence of a Gene-We next asked whether the stRNA-like silencing occurs in other genes and when targeted to the coding region of the mRNA. First, we designed a perfectly base-paired siRNA targeting the coding sequence of geminin mRNA (GEM2) (Fig. 2A). GEM2 siRNA reduces the protein and mRNA levels (Fig. 2B, quantitation in Fig. 2C). We then steadily incorporated mismatches within the siRNA. When three mismatches were incorporated with respect to the target (GEM3), there was significant decrease in protein but not in mRNA levels. Therefore, three mismatches in a geminin siRNA targeted to the coding region converts it to a function like an stRNA. Translational repression could be observed with three (GEM3) but not with five or eleven (GEM8, 7) mismatches (Fig. 2, A-C). Therefore stRNA-like silencing occurs on other endogenous genes and is not limited to the 3Ј-UTR.
Because detection of protein levels by enhanced chemiluminesence immunoblotting can be non-linear, we loaded varying amounts of cell lysate for immunoblotting to carefully quantitate the decrease in protein levels. Around 4-fold more cell lysate had to be loaded from GEM3-transfected cells as compared with control cells to detect equivalent geminin levels (Fig. 2D, quantitation in Fig. 2E). This decrease to ϳ25% of endogenous protein levels was comparable with the decrease seen with perfectly matched siRNA.
G⅐U Wobble Base Pairing in Antisense RNA Oligonucleotide-mRNA Target Interaction-The RNA secondary structure permits G⅐U wobble base pairs that have a thermodynamic stability comparable with Watson-Crick base pairs (37). If wobble base pairing is tolerated in the antisense RNA-target duplex, it would expand the range of potential targets for the same siRNA. To test whether wobble base pairing is tolerated in RNA-mediated silencing, we used a perfectly matched p21 siRNA that causes mRNA degradation (p21-2) (Fig. 3, A and  B). Three nucleotides were replaced such that the siRNA would form three G⅐U wobble base pairs with the target (p21-3). The protein silencing and the mRNA degradation were unaltered. On the other hand, replacing the same three nucleotides such that there are "real" mismatches with the target (p21-4) still silences protein expression, but this is via an stRNA-like pathway (decrease in protein without change in mRNA levels; Fig. 3, A and B, with quantitation in Fig.   3C). Therefore the three G⅐U wobble base pairs were recognized as authentic base pairs, and replacing them with nonwobble base pairs (which are recognized as mismatches) leads to stRNA-type activity.
We next investigated whether wobble base pairs can expand the spectrum of stRNA-like silencing. It has been reported that G⅐U base pairing is recognized as a mismatch in micro-RNAs leading to disruption of silencing (34). GEM6 is a geminin dsRNA duplex that silences endogenous geminin by translational repression (Fig. 3, D and E, with quantitation in Fig. 3F). This dsRNA duplex has three real mismatches and two G⅐U wobble base pairings with respect to the target. It does not degrade the target mRNA but silences the target protein expression. When the two G⅐U base pairs were changed to nonwobble mismatches (five real mismatches; GEM8), there was a significant reduction in silencing (Fig. 3, D-F). Therefore, as in mRNA degradation, wobble base pairing functions as authentic base pairing in translational repression, expanding the gamut of potential off-target genes that may be silenced by stRNA-like processes.
Base Pairing in the Center of the Target Region Is Critical for RNA-mediated Silencing-We tested whether mismatches at the ends are better tolerated than mismatches at the center of the RNA oligonucleotide-mRNA duplex. The antisense strand of a transfected dsRNA duplex is believed to be the functional strand in gene silencing. A perfectly base-paired RNA (p21-1) targeting the coding sequence of the p21 gene silences p21 efficiently (Fig. 4, A and B). When base pairing was disrupted at the 3Ј-end of the complementary region, we still observed effective silencing (p21-1g). However, when mismatches were introduced at the center of the base-pairing region (position 7-13 relative to the 5Ј-end of the antisense strand), no silencing was observed (p21-1h). Therefore, mismatches were better tolerated at the end of the duplex than in the middle. It is also of interest that a two-base pair mismatch in p21-1h abrogated silencing, whereas mismatches of three or four bases in p21-4, p21-10, and p21-11 still permitted silencing. Therefore, it is not simply the number of mismatches but also the context or nature of the mismatch that determines whether an oligonucleotide functions like an stRNA.
Translational Repression Can Be Induced by Single-stranded Antisense RNA-We next determined whether some of the rules by which siRNAs function in mammalian cells were also applicable to gene silencing via translational repression. Single-stranded antisense RNA has been shown to be effective as an siRNA (38). We therefore asked whether the single-stranded antisense RNA can induce silencing by translational repression. Consistent with published results, antisense single strand siRNA was as effective as double-stranded siRNA (GEM2) for silencing geminin expression (Fig. 4C, with quantitation in Fig.  4D). Transfecting the antisense strand of GEM3 also produced effective silencing, just like double-stranded GEM3 (Fig. 4, C  and D). The sense strand of GEM3 did not change geminin levels. Therefore, a single-stranded antisense RNA that can imperfectly pair with the target mRNA can induce translational repression.
5Ј-Phosphate Is Required for stRNA-like Silencing in HeLa Cells-The antisense strand of siRNA needs a 5Ј-phosphate group for activity (38,39). For the studies above, we used 5Ј-hydroxylated dsRNA, which is believed to be rapidly phosphorylated in vivo by endogenous kinases. We investigated the requirement of a 5Ј-phosphate group on the antisense strand of an RNA duplex for translational repression. For this we used an RNA duplex blocked at the 5Ј-end by an amino group with a three-carbon linker (as/5Ј-N3; Fig. 5A). These modified RNA duplexes lack the 5Ј-hydroxyl group and cannot be phosphorylated by kinases in vivo. We observed that these modified RNA duplexes did not silence gene expression (Fig. 5B, with quantitation in Fig. 5C). Therefore, similar to siRNA, the 5Ј-phosphate group on the antisense strand is essential for dsRNA to enter into the "stRNA-like silencing complex" to cause translational repression (38,39).
The 3Ј-hydroxyl group in the antisense strand of siRNA, which is essential in C. elegans, is dispensable for silencing in mammalian cells. We therefore tested whether stRNA-like silencing requires the 3Ј-hydroxyl group as a primer for RNA synthesis following annealing to the target RNA-template. The 3Ј-hydroxyl group of the antisense strand was blocked with 3-aminopropyl phosphoester (3Ј-amino modifier) so that it could no longer be used as primer for extension (as/3Ј-N3-3Ј; Fig. 5A). There was no difference in silencing with 3Ј-modified oligonucleotides, suggesting that RNA-mediated translational repression does not involve amplification of RNA by primer extension from the 3Ј-hydroxyl end (Fig. 5, B and C). Mismatches Are Tolerated at the 3Ј-End but Not the 5Ј-End of the Antisense Strand-In light of the observation that modifications at the 5Ј-end are not tolerated in RNA-mediated silencing, we tested whether mismatches at the 5Ј-end are tolerated. Unlike mismatches at the 3Ј-end, incorporating mismatches at the 5Ј-end of and antisense strand abolished silencing (Fig. 5, D and E, with quantitation in Fig. 5F). Therefore, for the p21 mRNA, incorporating Ͼ3-4 mismatches in the center of the target region or mismatches at the 5Ј-end of the antisense strand limits RNA-mediated silencing. DISCUSSION Different workers have reported varying degrees of tolerance for mismatches in siRNA-mediated silencing (13,(31)(32)(33). Al-though some report that a bulge or a 2-3 nucleotide mismatch is permissible in siRNA-mediated silencing, others observe that even a single base pair mismatch completely eliminates silencing (35,40). Here, we have systematically increased the mismatches between the short RNA and target RNA. We observe that, as mismatches increase, an siRNA begins to function like an stRNA. Therefore, we note that a 21-nucleotide double-stranded RNA that shares only partial complementarity with the target mRNA is still competent to cause gene silencing via translational repression. We also note that the antisense strand alone, but not the sense strand of the same duplex, can induce silencing.
Adaptation of an siRNA to function like an stRNA implies that complementarity is a critical factor defining the mode of action of these small RNA molecules. Hutvagner and Zamore demonstrated that in human cell extracts a natural stRNA, let-7, can function as siRNA with a perfectly base-paired target (41). Zeng et al. designed an miRNA that inhibited the translation of a target mRNA with which it shared partial complementarity (40). The target was an exogenously expressed mRNA in which four copies of the partial complementary sequence were inserted at the 3Ј-UTR. Similarly, Doench et al. reported that siRNA can function like an stRNA if it contains a central bulge with respect to the target (35). These workers again used an exogenously expressed mRNA in which four consecutive binding sites were engineered in the 3Ј-UTR. Here, we report that an RNA duplex can silence an endogenous gene with which it has a single binding site with up to 3-4 mismatches. This was demonstrated using two genes (p21 and geminin) in two different cell lines (PC3 and HeLa).
Various reports in literature suggest that the target of an stRNA has to be the 3Ј-UTR (22)(23)(24). However, recently Kawasaki and Taira have reported a natural micro-RNA, miR-23, that is partially complementary to the coding region and stop codon of the target mRNA (42). We now show that a mismatched RNA designed to target the coding sequence of the gene can function like an stRNA.
The perfectly base-paired siRNA (GEM2) silences the gene by mediating degradation of geminin mRNA. Although the silencing was complete, with undetectable protein levels, low levels of mRNA could still be detected. Therefore, it is possible that gene silencing mediated even by a perfectly base-paired siRNA like GEM2 involved a component of translational repression. It is likely that this serves as an additional or redundant mechanism (in case the mRNA target degradation is unsuccessful for some reason) to effect complete silencing.
We have identified the characteristics of dsRNA important for it to function like an stRNA. The 5Ј-phosphate but not the 3Ј-hydroxyl group of the antisense strand is required for translational repression in human cells. The requirement of the 5Ј-phosphate for stRNA-like activity is similar to that observed in siRNA (38,39). On the other hand, it is clear that, as with siRNA in human cells, amplification of an RNA duplex by primer extension from the 3Ј-hydroxyl end of the antisense strand is not required for translational repression in human cells.
Can we definitely predict the efficacy of dsRNA-mediated silencing based on the number of mismatches with the target? Two mismatches with the p21 mRNA (p21-h) abolishes silencing, whereas the same number of mismatches with geminin mRNA (GEM3) permits silencing. Therefore, it is not the absolute number of mismatches but probably the overall stability of the antisense RNA oligonucleotide-mRNA duplex that determines the success of a silencing event. This study suggests that additional experiments are necessary to explicitly predict the efficacy of silencing of different targets and sequences on the basis of number of mismatches. The secondary structure of the mRNA target and the sequence composition of the short RNA likely contribute to the stability of the duplex. Therefore, additional targets and sequences should be tested systematically to define broad rules for dsRNA-mediated silencing.
Is there any advantage for a cell to allow gene silencing by short RNA that has only partial complementarity? It has been suggested that si/stRNAs evolved as a defense mechanism against viruses and other intracellular pathogens. Because mismatched oligonucleotides can suppress protein expression, pathogens with a few nucleotide changes at the target site will still be inhibited by RNA duplexes. In addition, a single short RNA to a relatively conserved region might inhibit multiple cellular genes in the same family. It should also be noted that, because short RNA duplexes with mismatches and wobble base pairs can silence endogenous genes, programs searching for targets of siRNA should allow for such mismatches and wobble base pairs.
There are numerous reports in the literature describing the phenotypes following inactivation of various genes by siRNA duplexes (43). This study cautions that the silencing phenotypes observed by small RNA-mediated silencing should be interpreted conservatively, because there are now at least two additional mechanisms by which the target range is extended, namely mismatches and G⅐U wobble base pairs. Recently Jackson et al. have shown that partial complementarity between siRNA and mRNA leads to nonspecific changes in mRNA levels (29). We show here that partial complementarity can also lead to translational repression, thereby increasing the range of nonspecific effects. To eliminate off-target silencing effects, a simple precaution is to use two or more separate duplexes targeting different regions of an mRNA to look for a reproducible phenotype.