• Open Access

Selective silencing by RNAi of a dominant allele that causes amyotrophic lateral sclerosis


Zuoshang Xu, Department of Biochemistry and Molecular Pharmacology, University of Massachusetts Medical School, 364 Plantation St, Worcester, MA 01605–2324, USA. Tel.: +1 508 856 3309; fax: +1 508 856 8390; e-mail: zuoshang.xu@umassmed.edu
Phillip Zamore, as above. Tel.: +1 508 856 2191; fax: +1 508 856 2003; e-mail: phillip.zamore@umassmed.edu


RNA interference (RNAi) can achieve sequence-selective inactivation of gene expression in a wide variety of eukaryotes by introducing double-stranded RNA corresponding to the target gene. Here we explore the potential of RNAi as a therapy for amyotrophic lateral sclerosis (ALS) caused by mutations in the Cu, Zn superoxide dismutase (SOD1) gene. Although the mutant SOD1 is toxic, the wild-type SOD1 performs important functions. Therefore, the ideal therapeutic strategy should be to selectively inhibit the mutant, but not the wild-type SOD1 expression. Because most SOD1 mutations are single nucleotide changes, to selectively silence the mutant requires single-nucleotide specificity. By coupling rational design of small interfering RNAs (siRNAs) with their validation in RNAi reactions in vitro and in vivo, we have identified siRNA sequences with this specificity. A similarly designed sequence, when expressed as small hairpin RNA (shRNA) under the control of an RNA polymerase III (pol III) promoter, retains the single-nucleotide specificity. Thus, RNAi is a promising therapy for ALS and other disorders caused by dominant, gain-of-function gene mutations.


Amyotrophic lateral sclerosis (ALS) is a neurodegenerative disease that causes motor neuron degeneration, skeletal muscle atrophy and paralysis. This disease is progressive and invariably fatal, resulting in the death of the patient within 1–5 years after diagnosis. At present there is no cure (Rowland & Shneider, 2001). A fraction of ALS is caused by mutations in the Cu, Zn superoxide dismutase (SOD1) gene (Rosen et al., 1993). These mutations cause motor neuron degeneration because the mutant protein has acquired some toxic property (Cleveland & Rothstein, 2001). Neither the molecular basis of this toxic property nor the way in which the toxic protein triggers motor neuron degeneration is understood. In mice, expression of mutant SOD1, but not complete elimination of SOD1, causes ALS. Nonetheless, SOD1-knockout mice show reduced fertility (Matzuk et al., 1998), motor axonopathy (Shefner et al., 1999), age-associated loss of cochlear hair cells (McFadden et al., 2001) and neuromuscular junction synapses (Flood et al., 1999), as well as enhanced susceptibility to a variety of noxious assaults on the nervous system, such as axonal injury (Reaume et al., 1996), ischaemia (Kondo et al., 1997; Kawase et al., 1999), haemolysate exposure (Matz et al., 2000) and irradiation (Behndig et al., 2001). Given the toxicity of the mutant protein and the functional importance of the wild-type, the ideal therapy for ALS would selectively block expression of the mutant while retaining expression of wild-type protein.

Sequence-selective inactivation of gene expression can be achieved in a wide variety of eukaryotes by introducing double-stranded RNA corresponding to the target gene, a phenomenon termed RNA interference (RNAi) (Hannon, 2002; Hutvagner & Zamore, 2002; McManus & Sharp, 2002). Early mechanistic studies of the RNAi pathway (Hamilton & Baulcombe, 1999; Hammond et al., 2000; Zamore et al., 2000; Elbashir et al., 2001b) led to its extension to cultured mammalian cells (Caplen et al., 2001, Elbashir et al., 2001a). In this approach, small interfering RNAs (siRNAs), 21 nt, double-stranded RNA intermediates in the RNAi pathway, are used to trigger the destruction of an mRNA containing the siRNA sequence. More recent experiments indicate that short hairpin RNAs (shRNAs) transcribed in vivo can trigger degradation of a corresponding mRNA, because shRNA is processed into siRNA in cells (Brummelkamp et al., 2002b; Jacque et al., 2002; Lee et al., 2002; McManus et al., 2002; Miyagishi & Taira, 2002; Paddison et al., 2002; Paul et al., 2002; Sui et al., 2002; Yu et al., 2002; Zeng et al., 2002). These technical advances raise the possibility that siRNA duplexes or viruses expressing shRNA may be used to block the expression of a mutant gene.

Because the vast majority of ALS-causing SOD1 mutations are single-nucleotide point mutations that alter a single amino acid in the protein (http://www.alsod.org/), the first step in developing RNAi therapy is to identify siRNA and shRNA sequences that can selectively silence the expression of mutant, but not wild-type, protein with single-nucleotide specificity. RNAi is a promising strategy for allele-specific silencing, but the design of siRNAs with single-nucleotide specificity is not straightforward. siRNAs that differ from the sequence of their target RNA at one or more nucleotides retain efficacy in some cases (Boutla et al., 2001, Holen et al., 2002) and lose activity in others (Boutla et al., 2001; Elbashir et al., 2001c; Brummelkamp et al., 2002a,b, Yu et al., 2002). Here we coupled rational design with validation in RNAi reactions in vitro and in vivo, and developed siRNA and shRNA sequences that selectively silence two mutant SOD1 alleles but not the wild-type. These sequences can be developed further to treat ALS caused by these SOD1 mutations.


Mechanistic studies suggest that formation of an A-form helix between the siRNA and its mRNA target is required for mRNA cleavage (Chiu & Rana, 2002). We reasoned that mismatches at or near the site of target cleavage would disrupt the required A-form helix. We targeted an allele of SOD1 in which guanosine 256 (G256, relative to the start of translation) is mutated to cytosine, generating a glycine-to-arginine mutation (G85R). We placed the mismatch at positions 9, 10 and 11 from the 5′ end of the siRNA. The G256C mutant/wild-type pair produces the largest possible clash (purine:purine) between the mutant siRNA and the wild-type gene and the greatest hydrogen-bonding (G:C) between the mutant siRNA and the mutant SOD1 allele (Fig. 1A). As controls, we synthesized comparable siRNAs to target wild-type but not mutant SOD1 mRNA (Fig. 1A). In the controls, the siRNAs contain a G:C base pair at the selective site, but the mismatch between wild-type siRNA and the mutant allele is a smaller pyrimidine:pyrimidine clash (C:C). The selectivity of each siRNA was tested in a cell-free RNAi reaction containing Drosophila embryo lysate (Tuschl et al., 1999; Zamore et al., 2000) (Fig. 1B,C).

Figure 1.

siRNA duplexes can discriminate between mutant and wild-type SOD1 in vitro. (A) siRNA duplexes used. (B) In vitro RNAi experiments targeting mutant or wild-type SOD1 mRNA with mutant or wild-type siRNAs. (C) Mutant siRNA p10 targets mutant (red curves) but not wild-type SOD1 (blue curves) mRNA for destruction by the RNAi pathway.

Each of the six siRNAs cleaved the corresponding target RNA, although with dramatically different efficiency (Fig. 1B). For example, neither mutant nor wild-type p11 siRNAs cut their respective RNA targets with a rate expected to be effective in vivo. By contrast, the p10 mutant siRNA efficiently cleaved the mutant SOD1 mRNA. In all cases, destruction of full-length target mRNA was accompanied by a corresponding accumulation of a 300 nt 5′ cleavage product, demonstrating that the siRNAs trigger RNAi, rather than non-specific RNA degradation (Fig. 1B). In the absence of siRNA, or in the presence of an unrelated siRNA, the mutant SOD1 target RNA was stable in the Drosophila embryo lysate (data not shown). Data for both the destruction of target RNA and the accumulation of 5′ cleavage product fit well to a single exponential equation, indicating that the reaction follows pseudo first-order kinetics (Fig. 1C).

To determine the selectivity of the six siRNAs, each siRNA corresponding to the mutant SOD1 sequence was tested for its ability to cleave wild-type SOD1 mRNA, and each wild-type siRNA was tested for its ability to cleave mutant mRNA. Some but not all of the siRNA duplexes effectively discriminated between the target to which they were perfectly matched and the target with which they had a single-nucleotide mismatch (Fig. 1B). We observed two types of defects for a subset of siRNAs. Both wild-type and mutant p11 siRNA did not trigger efficient target cleavage of either the perfectly matched or the mismatched RNA target (Fig. 1B). Thus, these siRNA sequences are inherently poor triggers of RNAi. The p9 and p10 wild-type siRNAs not only triggered rapid cleavage of their corresponding wild-type target, but also produced significant cleavage of the mutant RNA (Fig. 1B). These siRNAs are good triggers of RNAi but show poor selectivity. By contrast, the p10 mutant siRNA showed both efficient RNAi and robust discrimination between mutant and wild-type SOD1 RNAs, cleaving the mutant far more efficiently than the wild-type RNA in the cell-free reaction (Fig. 1B,C). Because this siRNA showed nearly complete discrimination between mutant and wild-type SOD1 mRNA targets (Fig. 1B,C), it is an ideal candidate for therapeutic application.

To test whether the cell-free reactions accurately predict siRNA efficacy and selectivity in mammalian cells, we analysed the siRNAs in a HeLa cell assay. We prepared plasmids that expressed either SOD1WT or SOD1G85R with GFP fused to their carboxyl terminus. Each construct was transfected into HeLa cells together with both siRNA and a dsRed-expressing vector that served as a transfection control. The expression of either mutant or wild-type SOD1 was monitored by fluorescence-activated cell sorting (FACS). Transfection of p9, p10 and p11 siRNAs with their corresponding mutant or wild-type targets suppressed gene expression, although with distinctly different efficiency and selectivity (Fig. 2). Co-transfection with an siRNA complementary to firefly luciferase did not suppress either SOD1 allele (Fig. 2). As observed in the cell-free reactions, the p10 siRNA against wild-type SOD1 showed no selectivity and suppressed both wild-type and mutant SOD1 mRNA (Fig. 2). The other siRNAs all showed some degree of selectivity, but the p10 siRNA directed against the SOD1 mutant mRNA showed both the greatest efficacy and selectivity, in agreement with the results of the cell-free reactions. Thus, some but not all siRNAs can efficiently discriminate between mRNA targets with a single-nucleotide difference.

Figure 2.

Selective inhibition of mutant SOD1G85R expression by siRNA in HeLa cells. SOD1WT–GFP or SOD1G85R–GFP was co-transfected with various siRNAs. DsRed was co-transfected as a transfection control. Green and red fluorescent cells were quantified using FACS. (A) Raw FACS cell counts; (B) relative number of green and red cells in the transfections (n = 3). Error bars represent standard deviation.

Recently, it has been shown that shRNA can trigger RNAi in vivo. To test whether shRNA against mutant SOD1 can selectively block the expression of the mutant, but not the wild-type, SOD1 expression, we constructed a plasmid that synthesizes an shRNA homologous to another disease-causing mutant SOD1G93A (Sui et al., 2002). This mutant was examined because, like the SOD1G85R, it is a guanosine to cytosine change at neucleotide position 281, thus placing a G:G mismatch at the selective site between the shRNA and wild-type SOD1 (Fig. 3A). When co-transfected separately with mutant SOD1–GFP, or wild-type expressing plasmid, this hairpin construct inhibited mutant, but not wild-type, expression (Fig. 3). Thus, hairpin constructs can be used to trigger single-nucleotide selective RNAi of mutant SOD1 in cultured cells. To test if mutant-selective inhibition can be achieved in neuronal cells, we separately transfected the wild-type or mutant SOD1–GFP constructs with either siRNA p10 against SOD1G85R or shRNA-synthesizing vector against SOD1G93A into the neuroblastoma cell line N2a. As in HeLa cells, both synthetic siRNAs and shRNA constructs directed the selective inhibition of mutant SOD1 expression in N2a cells (Fig. 4A,B).

Figure 3.

Selective inhibition of mutant SOD1G93A expression by U6-G93A vector in HeLa cells. (A) Design of the G93A shRNA. Bold letters mark nucleotides that match the SOD1 sequence. The underlined letter marks the mismatched nucleotide. (B) SOD1WT–GFP or SOD1G93A–GFP was co-transfected with U6-empty or U6-G93A (U6 constructs : SOD1–GFP = 5 : 1). DsRed was co-transfected as a transfection control (SOD1–GFP : DsRed= 4 : 1). Green and red fluorescent cells were quantified using FACS. Results from four experiments were averaged. Error bars represent standard deviation.

Figure 4.

Selective inhibition of mutant SOD1 expression by siRNA and U6-G93A vector in neuroblastoma N2a cells. Green and red fluorescent cells were quantified using FACS. (A) siRNA against SOD1G85R (n = 4); (B) U6-G93A vector (n = 3). Error bars represent standard deviation.

To be therapeutically relevant, single-nucleotide selective siRNAs must discriminate between mutant and wild-type SOD1 when both mRNAs are present in the same cell. We transfected HeLa cells with p10 siRNAs and mutant SOD1G85R–GFP, and analysed SOD1 protein expression by immunoblotting with anti-SOD1 antibody that recognizes both the transfected SOD1G85R–GFP fusion protein and endogenous wild-type SOD1. As expected, p10 siRNA against wild-type SOD1 inhibited both the endogenous wild-type SOD1 and the transfected SOD1G85R–GFP (Fig. 5). The near 50% inhibition of the endogenous wild-type SOD1 expression probably reflected the transfection efficiency, which was ∼50%. In contrast to the p10 wild-type siRNA, at two different doses, p10 siRNA against the mutant inhibited expression of the mutant, but had no effect on the expression of endogenous wild-type SOD1 (Fig. 5). No additional selectivity was seen with a 3′-blocked siRNA, consistent with reports that siRNAs do not function as primers to trigger the production of ‘secondary siRNAs’ in human cells (Chiu & Rana, 2002; Holen et al., 2002; Schwarz et al., 2002), as they do in nematodes (Sijen et al., 2001).

Figure 5.

Selective inhibition of mutant SOD1 G85R but not the wild-type SOD1 expression by siRNA in the same cells. (A) Protein blots of the transfected HeLa cells detecting mutant SOD1G85R–GFP and endogenous human SOD1. (B) Relative levels of SOD1 measured from the protein blots (average of four transfections). Error bars are standard error.

Finally, to test whether this selective inhibition can occur in vivo, we transfected SOD1 reporters and shRNA plasmid into mice using the hydrodynamic transfection protocol. Mutant SOD1G93A–GFP and myc-tagged wild-type human SOD1 expression plasmids were used, enabling detection of mutant and wild-type human SOD1 proteins, as well as the endogenous mouse SOD1 by immunoblotting. We analysed SOD1 expression in liver, a tissue readily transfectable by the hydrodynamic method. Under these conditions, the shRNA-expressing plasmid selectively decreased mutant, but not wild-type, human SOD1 expression (Fig. 6).

Figure 6.

Selective inhibition of mutant SOD1 expression by U6-G93A vector in vivo. (A) SOD1G93A–GFP was co-transfected with a C-terminal myc-tagged wild-type human SOD1 in mice using the hydrodynamic transfection method. Liver proteins were resolved on SDS-PAGE and detected using anti-SOD1 antibody. (B) The relative band intensities were quantified. The ratio of SOD1G93A–GFP to wild-type SOD1 myc is shown. Eight animals were used in each group. The U6-G93A group is significantly different from the other two groups (P < 0.05) as determined from Hohm's modified Bonferroni procedure for multiple t-tests.


We have found siRNA and shRNA sequences that selectively silence two dominant mutant SOD1 genes. Using multiple siRNAs matching either wild-type or mutant SOD1, we show that a subset of siRNAs against mutant SOD1 cleave the mutant, but not the wild-type SOD1 RNA efficiently in vitro (Fig. 1). Those siRNAs showing both efficacy and selectivity in vitro also selectively inhibit mutant but not wild-type SOD1 protein expression in mammalian cells (Figs 2 and 4), even when both the mutant and the wild-type proteins are present in the same cells (Fig. 5). Furthermore, a vector expressing shRNA similarly designed according to the optimal siRNA also selectively inhibited mutant but not wild-type SOD1 expression in mouse liver (Figs 3, 4 and 6). These results demonstrate that selective inhibition of dominant mutant SOD1 alleles can be achieved using RNAi and the optimal siRNA and shRNA sequences can be identified by a preclinical screen in vitro and in vivo.

Our search for siRNA sequences optimized for selective silencing of the mutant but not the wild-type SOD1 revealed that single nucleotide discrimination is not guaranteed. Some siRNAs can discriminate between alleles that differ at a single nucleotide, whereas others cannot. Our results point to two different types of deficiencies for siRNAs designed to target mutant, disease-causing alleles. First, not all siRNAs silence with the same efficiency. Among the siRNAs directed against wild-type SOD1, p9 and p10 cleave the target more efficiently than p11 (Fig. 1). As predicted by analysis in cell-free RNAi reaction, the p10 siRNA inhibited target gene expression most efficiently in mammalian cells (Fig. 2). Among the siRNAs against the mutant SOD1G85R, p9 and p10 cleave the mutant far more than p11 (Fig. 1). As in the cell-free assay, p10 was the most efficient siRNA in inhibiting the mutant SOD1 expression in mammalian cells (Fig. 2). It is intriguing that displacing the siRNA along the target sequence by a single nucleotide results in such a dramatic change in silencing efficiency. Second, significant differences in selectivity between the perfectly matched target RNA and the RNA bearing a single-nucleotide mismatch were observed among the six siRNAs used. For example, wild-type p10 siRNA had poor selectivity; it cleaved both wild-type and mutant SOD1 RNA in the cell-free assay and efficiently inhibited the expression of both alleles in mammalian cells (Figs 1, 2, 4 and 5). By contrast, the p10 siRNA directed against mutant SOD1 showed nearly perfect selectivity. It cleaved mutant SOD1 RNA but not wild-type, in the cell-free assay, and inhibited mutant but not wild-type SOD1 expression in mammalian cells (Figs 1, 2, 4 and 5).

Our results raise questions regarding the rules in designing optimal siRNA or shRNA for single nucleotide discrimination. Among the contributing factors is the type of mismatch at the critical site p10. We predict that a purine:purine mismatch disrupts the A-form helix that is required between the antisense strand of the siRNA and its mRNA target (Chiu & Rana, 2002). By contrast, a pyrimidine:pyrimidine mismatch may more readily be accommodated within an A-form helix. Thus, the G:G clash between the siRNA and the wild-type target RNA discriminates against the wild-type target, producing greater selectivity for the mutant target, whereas the presence of a G:C basepair p79 between the mutant siRNA and the mutant target mRNA at the selective site may serve to maximize the energy difference between mismatch and perfect pairing (see Figs 1A and 3A). Consistent with this view, in Drosophila embryos, an siRNA having a pyrimidine:purine mismatch (C:A) with its target mRNA was only slightly less effective than the perfectly matched siRNA (Boutla et al., 2001). Moreover, an siRNA directed against firefly luciferase failed to produce detectable RNAi in vitro when it contained at position 9 or 10 of its guide strand a purine:purine (A:A) mismatch with its target RNA (Elbashir et al., 2001c). Similarly, an siRNA that showed good selectivity for a mutant Ras mRNA created a purine:purine (A:G) clash with the wild-type allele (Brummelkamp et al., 2002a). Arguing against this view, one experiment using siRNA against hTF suggests that a G:G mismatch can still mediate RNAi, albeit with reduced efficiency (Holen et al., 2002). It is possible that this was due to the high concentration of siRNA used. Another experiment using shRNA against CDH-1 suggests that a U:C or a U:G mismatch abolished RNAi (Boutla et al., 2001). In light of our demonstration that small differences in siRNA sequence can produce dramatic differences in efficacy, rather than selectivity, it remains to be shown if these inactive shRNAs were active against a perfectly matched target and not merely poor triggers of RNAi in general. Clearly, further work is required to clarify the rules in designing siRNA and shRNA sequences optimized for selective silencing of mutant alleles with single-nucleotide specificity.

Taken together, we have identified siRNA and shRNA sequences that can selectively down-regulate the expression of mutant but not the wild-type SOD1, even when the mutant mRNA differs from wild-type by a single nucleotide. Because the shRNA-synthesizing plasmid construct can be readily incorporated into viral vectors (Brummelkamp et al., 2002a; Devroe & Silver, 2002; Xia et al., 2002), these siRNA and shRNA sequences can be readily placed in virus-based delivery systems to treat ALS caused by mutant SOD1 expression. In broad terms, our results show the promise of RNAi as a therapeutic strategy to diseases caused by dominant, gain-of-function gene mutations.

Experimental procedures

RNA and DNA constructs

Twenty-one nucleotide single-strand RNAs (Fig. 1) were purchased from Dharmacon Research, deprotected according to the manufacturer's instructions, and annealed as described (Nykanen et al., 2001). The 3′-block siRNA was synthesized as 2′,3′-dideoxy cytidine at the 3′ terminus of the antisense strand. To create wild-type and mutant SOD1–GFP fusion proteins, SOD1WT, SOD1G85R and SOD1G93A cDNAs (kind gifts of Dr Joseph Beckman) were PCR cloned between the PmlI and PstI sites of pCMV/myc/mito/GFP (Invitrogen). This cloning step deleted the mitochondrial targeting sequence. To create myc-tagged wild-type SOD1, SOD1wt cDNA was PCR cloned between the PstI and XhoI sites of pCMV/myc/mito/GFP. The mitochondrial targeting sequence was then deleted by digestion with BssHII and PmlI and blunt ligation. All constructs were verified by sequencing. DsRed (pDsRed2-C1) was purchased from Clontech (Palo Alto, CA, USA). U6-G93A was constructed as described (Sui et al., 2002) (Fig. 3).

In vitro RNAi assay

Drosophila embryo lysates were prepared as previously described (Tuschl et al., 1999, Zamore et al., 2000). Five hundred and sixty nucleotide human SOD1 target RNAs containing either wild-type or mutant SOD1G85R coding sequence were cap-labelled using guanylyl transferase as described previously (Zamore et al., 2000). In vitro RNAi reactions were carried out in Drosophila embryo lysate by incubating ∼5 nm of the 5′, [32P]cap-radiolabelled target RNA with 100 nm duplex siRNA at 25 °C in a standard reaction (Tuschl et al., 1999; Zamore et al., 2000). Cleavage products were analysed on 5% denaturing acrylamide gels, dried and exposed on image plates (Fuji). Plates were scanned using a Molecular Imager FX (BioRad), and images were analysed using Quantity One version 4.0.3 (BioRad).

Cell culture and transfection

HeLa cells were cultured in DMEM and N2A cells in DMED and Opti-MEM (1 : 1), both supplemented with 10% fetal bovine serum (FBS), 100 u mL−1 penicillin and 100 µg mL−1 streptomycin. Twenty-four hours before transfection, cells (70–90% confluency) were detached by trituration, transferred to six-well plates and cultured in fresh medium without antibiotics. Transfection was carried out using lipofectamine 2000 (invitrogen) according to the manufacturer's instructions. The amounts of the constructs used in transfections are 4 µg each of mutant or wild-type SOD1–GFP and DsRed plasmids, 4 × 10−11 or 4 × 10−12 mol siRNAs, and 20 or 8 µg U6-G93A, unless stated otherwise.

In vivo transfection

Twenty-four 6–8-week-old mice were divided into three groups. The first group received no shRNA vector, the second group received 20 µg empty vector, and the third group received 20 µg U6-shRNA vector against SOD1G93A. All groups received both 20 µg of myc-tagged human wild-type SOD1 and 20 µg GFP-tagged SOD1. The vectors were diluted in Ringer's solution so that the total volume was 2.5 mL per mouse. Mice were anaesthetized with avertin (240 mg kg−1) and the vectors were injected into the tail vein using a 26-gauge needle in less than 10 s. Forty-eight hours following injection, animals were perfused with 5 mL PBS in order to remove blood from the liver. Livers were dissected and quickly frozen on dry ice. Samples were placed in 25 mm PBS buffer (pH 7.2) containing 1% SDS, 1 mm DTT, 1 mm phenylmethylsuphonyl fluoride (PMSF), and protease inhibitor cocktail (Sigma, diluted 1 : 100) and homogenized using a hand-held polytrone (Pro-scientific).

Western blot analysis

Protein concentrations were determined using a BCA protein assay kit (Pierce, Rockville, IL, USA). Twenty-five micrograms HeLa cell proteins or 100 µg liver proteins were separated on a 15% SDS–PAGE gel and transferred onto Genescreen Plus membrane (Perkin Elmer). Rabbit anti-SOD1 (Biodesign) or Sheep anti-SOD1 was the primary, and HRP-labelled goat anti-rabbit IgG (Amersham) or donkey anti-sheep IgG was the secondary antibody. The protein bands were visualized using SuperSignal kit (Pierce) and Kodak Digital Image Station 440CF. The intensity of the bands was quantified using Kodak 1D software.


This work was supported by grants from the NINDS (NS35750), ALS Association and Robert Packard Center for ALS Research at Johns Hopkins to Z.S.X., from the NIH to P.D.Z. (GM62862) and to Y.S. (GM53874). P.D.Z. is a Pew Scholar in the Biomedical Sciences and a W. M. Keck Foundation Young Scholar in Medical Research. We thank Ms Ellen Trang for animal maintenance, Dr H. Tummala for critically reading the manuscript, and Drs Cindy Higgins, Xugang Xia and members of the Zamore Laboratories for advice and support.