REVIEWS

IMPROVING THE EFFICIENCY OF RNA INTERFERENCE IN MAMMALS Vivek Mittal RNA interference (RNAi) has been very successfully applied as a gene-silencing technology in both plants and invertebrates, but many practical obstacles need to be overcome before it becomes viable in mammalian systems. Greater specificity and efficiency of RNAi in mammals is being achieved by improving the design and selection of small interfering RNAs (siRNAs), by increasing the efficacy of their delivery to cells and organisms, and by engineering their conditional expression. Genome-wide functional RNAi screens, which are predominantly done in worms and flies, have now begun to revolutionize large-scale loss-of-function studies in mammals.

INTERFERONS

A group of glycoprotein cytokines that are produced by animal cells when they are invaded by viruses. They are also activated by long dsRNA, which results in the activation of 2′–5′ oligoadenylate synthase and protein kinase PKR. The 2′–5′ oligonucleotide activates RNAseL, which leads to mRNA cleavage. PKR phosphorylates the translation initiation factor eI2α, which leads to global inhibition of mRNA translation.

Cancer Genome Research Center, Cold Spring Harbor Laboratory, 500 Sunnyside Boulevard, Woodbury, New York 11797, USA. e-mail: mittal@cshl. org doi:10.1038/nrg1323

NATURE REVIEWS | GENETICS

RNA interference (RNAi) is the process by which dsRNA silences gene expression, either by inducing the sequence-specific degradation of complementary mRNA or by inhibiting translation. In principle, any gene can be silenced, so RNAi provides a rapid means for assessing the effects of the loss of gene function. Indeed, genome-wide RNAi screens have been successfully carried out in plants and invertebrates (for reviews, see REFS 1–5). An advantage of these RNAi screens is that a loss-of-function phenotype can be readily linked to a specific gene; this compares favourably with mutagenesis screens, which require the laborious mapping of the induced mutation. Although gene-specific, long dsRNA molecules rapidly came into common use for targeted gene ‘knockdown’ in Caenorhabditis elegans and other organisms (for reviews, see REFS 1–3), it was initially thought that the approach would not be useful in mammals. This is because dsRNA molecules that are longer than 30 bp provoke the antiviral/INTERFERON PATHWAY, which results in the global shutdown of protein synthesis4. This view changed in 2001, with the crucial insight of Tuschl and colleagues6, and Caplen and colleagues7. They showed that chemically synthesized short dsRNA molecules of 21–22 nucleotides (nt) — known as small interfering RNAs (siRNAs) — could be used to target mammalian genes by RNAi while evading the interferon response6. Soon after, another important technical advance came from the demonstration that the endogenous expression

of siRNA in the form of short hairpin RNAs (shRNAs), which bear a fold-back stem-loop structure8–11, induced target-gene silencing in mammalian cells. Although these shorter dsRNAs usually bypass the interferon response, more recent reports indicate that this might not always be the case12,13,126, which indicates that improved siRNA designs and delivery systems are required to overcome this nonspecific effect. The ability of RNAi to efficiently suppress target genes in mammals has been heralded as the most powerful reverse-genetics approach in recent times. But can the power be exploited for the systematic inactivation of a large number of genes, to harness the dormant potential of sequenced mammalian genomes? For RNAi to be widely accepted as a gene-silencing tool in mammals significant technical challenges will need to be overcome. These challenges include developing efficient ways of designing, identifying and delivering effective siRNAs: specifically, improving the accuracy with which susceptible sites in the target RNA molecules can be identified, and obtaining inducible tissue- and cell-specific regulation of siRNA both in vitro and in vivo. In this article, I discuss how some of these technical challenges are being met and review the important developments that will allow functional gene-knockdown studies to be carried out successfully in mammalian systems. Insights from these studies are already allowing large-scale phenotypic screens to be carried out in mammals with unprecedented speed. These developments will probably also VOLUME 5 | MAY 2004 | 3 5 5

REVIEWS

Long dsRNA 5′-HO

OH-3′

3′-HO

OH-5′

Synthetic ds siRNA

Plasmid-based shRNA Dicer

ATP

ATP Endogenous miRNA ADP + Pi

ADP + Pi

5′-P

OH-3′

3′-HO

P-5′

siRNA ATP RISC ADP + Pi

siRNA function

miRNA function

5′-P

RISC 5′ ORF 3′-HO

RISC AAA P-5′

3′UTR RISC incorporates partially complementary mRNA

3′-HO

5′-P

RISC

3′-HO

RISC incorporates sense strand

OH-3′

RISC

OH-3′

P-5′

RISC incorporates antisense strand

RISC

P-5′

RISC incorporates sense and antisense strand

mRNA targeting

Translational repression

Specific or nonspecific gene silencing depending on the degree of homology

Figure 1 | RNAi-mediated gene silencing in mammals. The processing of long dsRNAs, hairpin microRNAs (miRNAs) or plasmid-synthesized short hairpin RNAs (shRNAs) by Dicer (an RNAse III family member) leads to the formation of small interfering RNAs (siRNAs) — 21–23-nucleotide (nt) RNA duplexes with symmetric 2–3-nt 3′ overhangs and 5′- phosphate groups. Exogenously provided synthetic siRNAs are converted into active functional siRNAs by an endogenous kinase that provides 5′-phosphate groups in the presence of adenosine triphosphate (ATP). siRNAs associate with cellular proteins to form an RNA-induced silencing complex (RISC), which contains a helicase that unwinds the duplex siRNA in an ATP-dependent reaction. In an ideal situation the antisense strand guides RISC to the target mRNA for endonucleolytic cleavage. In theory, each of the siRNA strands can be incorporated into RISC and direct RNA interference (RNAi). The antisense strand of an siRNA can direct the cleavage of a corresponding sense RNA target, whereas the sense strand of an siRNA can direct the cleavage of an antisense target. An RNA with a perfect match to a target mRNA behaves like an siRNA and results in mRNA degradation, whereas an RNA with a partial match functions as an miRNA and causes translational repression. Interestingly, recent data show that miRNAs can induce the degradation of fully complementary mRNA targets24. ORF, open reading frame.

HOMOLOGOUS RECOMBINATION

The process of replacing an endogenous gene with an artifical cassette that is flanked by regions that are homologous to sequences bordering the targeted gene. It is used as a technique to inactivate a gene and determine its function in a living animal. ANTISENSE VECTORS

Vectors that express DNA or RNA molecules that are complementary to sequences on target mRNA and result in the inhibition of protein synthesis.

356

| MAY 2004 | VOLUME 5

influence advances in siRNA-based therapeutics (this topic has been reviewed elsewhere14–16 and will not be described here further). RNAi is by no means the only method that is available for the targeted inhibition of mammalian gene expression. Other approaches, including traditional gene targeting by HOMOLOGOUS RECOMBINATION17, ANTISENSE VECTORS, catalytic RNA molecules (ribozymes) and catalytic DNA molecules (DNAzymes) (for a recent review, see REF. 18), have been successfully applied in some cases but, unlike RNAi, their cross-species application has been limited19. In addition, because RNAi is an endogenous natural pathway its power as a gene-silencing tool is likely to be orders of magnitude greater than other approaches.

The mechanism of gene silencing by RNAi

Accumulating biochemical and genetic evidence has begun to provide a mechanistic understanding of RNAimediated gene silencing in mammalian cells (FIG. 1) (for recent reviews, see REFS 3–5). Depending on the organism, RNAi is triggered by various types of molecule, including long dsRNAs, plasmid-based short hairpin RNAs (shRNAs) or endogenous hairpin micro RNAs (miRNAs). These are processed by the ribonuclease-III activity of the evolutionarily conserved Dicer enzyme6,20 to generate 21–22-nt siRNAs. These siRNAs are then incorporated into a protein complex, known as the RNA-induced silencing complex (RISC), which in turn uses an ATP-dependent RNA-helicase activity to

www.nature.com/reviews/genetics

REVIEWS unwind the duplex siRNA into single-stranded siRNA21. The antisense strand of the duplex siRNA guides the RISC to the homologous mRNA, where the RISC-associated endoribonuclease cleaves the target mRNA at a single site in the centre, which results in the silencing of the target gene6,22. Interestingly, miRNAs suppress the expression of partially complementary target mRNAs by translation inhibition rather than mRNA degradation (for recent reviews, see REFS 5,23). siRNA, shRNA and miRNA elicit RNAi through common biochemical pathways24. Generating effective siRNA probes

In mammalian systems, siRNA-based RNAi is rapidly emerging as an essential gene-silencing tool. However,

Box 1 | Designing siRNAs for gene silencing

Empirical rules for the design of siRNAs Although gene-specific long dsRNAs can efficiently induce gene silencing in invertebrates, RNA interference (RNAi) in mammalian somatic cells requires the identification of individual small interfering RNAs (siRNAs). On the basis of analyses of a small number of target genes that were successfully silenced, a set of empirical guidelines have been proposed for the design of siRNAs. These rules require the generation of two 21-nucleotide (nt) sense and antisense oligoribonucleotides that target a region in the gene to be silenced. It is important that each strand has 2-nt 3′-end overhangs. An example siRNA would be: sense 5′-(N19)TT-3′ and antisense 5′-(N19)TT-3′, where N is any nucleotide. The use of 2′-deoxythymidines for the 2-nt 3′ TTs has also been suggested as a way of safeguarding siRNAs from exonuclease activity. As an example, the design of an siRNA that targets the mammalian lamin A/C gene (LMNA)114 is shown below. Target region in LMNA mRNA: 5′…AACTGGACTTCCAGAAGAACATC…3′ Sense siRNA: 5′-CUGGACUUCCAGAAGAACAdTdT-3′ Antisense siRNA: 5′-UGUUCUUCUGGAAGUCCAGdTdT-3′ siRNA: Anneal Sense: 5′-CUGGACUUCCAGAAGAACAdTdT-3′ Antisense: 3′-dTdTGACCUGAAGGUCUUCUUGU-5′

Other suggestions include that the siRNA must avoid targeting certain regions of the mRNA: introns, both 5′ and 3′ untranslated regions (UTRs), regions within 75 bases of the start codon and sequences with >50% G+C content. Finally, to minimize nonspecific effects, a BLAST-search of genome sequence databases (NCBI Unigene and EST libraries, Celera databases; see online links box) should be carried out to ensure that only one gene is being targeted.

Limitations of empirical guidelines These guidelines have allowed the identification of siRNAs that promote efficient gene silencing; however, the rules have not been tested systematically on a large set of genes, and in many cases most siRNAs or shRNAs that are designed against a gene have not been effective32–34,37,51,115,116. The instability of the siRNA probe in vivo, its inability to interact with components of the RNAi machinery, and the inaccessibility of the target mRNA owing to local secondary structural constraints are all possible causes for the failure of most tested siRNAs. One proposed solution to the poor effectiveness of many siRNAs has been to deliver a mixture of siRNAs that are generated by using either Escherichia coli RNAse III (REFS. 117,118) or recombinant human Dicer33,119 to hydrolyse long dsRNAs. As the mixture contains many different siRNAs, gene silencing is virtually guaranteed. However, an important concern with such approaches is the possibility that nonspecific gene silencing will occur owing to the presence of several siRNA molecules, as well as the low effective concentration and lack of sequence information on the most potent siRNA in the mixture. Contrary to what is proposed in the guidelines, successful gene inhibition has been reported for siRNAs that target the UTRs of genes such as those encoding the Myc oncogene120 and the CD8a receptor51.

NATURE REVIEWS | GENETICS

the inability to design potent siRNAs to reliably target cognate mRNAs has imposed serious limitations on its general applicability. This has stimulated a drive to generate effective siRNAs that specifically silence every mammalian gene. Several groups have proposed basic empirical guidelines for designing effective siRNAs8,25; however, these rules should be regarded as a general guide and do not ensure that each selected siRNA will perform well (BOX 1). To overcome these problems, rational siRNA design schemes are being developed that are based on an understanding of RNAi biochemistry26–31 and on naturally occurring miRNA function. Although the rational designs have so far been encouraging in generating effective siRNAs, there are still many hurdles to be overcome in accurately predicting the siRNA that will be most effective in gene silencing. For example, little is known about which region in the target mRNA is most accessible to the siRNA. For this purpose, other groups are developing strategies for experimentally screeningout effective siRNAs from a panel of existing siRNAs that target many potential sites within a given mRNA32. The screening approach is likely to remain useful until we are able to predict one potent siRNA per gene with maximum accuracy. Rational design. The rational design of effective siRNA has paralleled rapid advances in two areas: our understanding of the biochemical mechanism of RNAi-mediated gene silencing and of the structure–function analysis of endogenous miRNAs. The crucial observation that each RISC contains only one of the two strands of the siRNA duplex22, and that the antisense strands of an siRNA can only direct the cleavage of the sense mRNA target has provided important insights into the design of effective siRNA26–31. Zamore and colleagues27 showed that the 5′ end of the strand that was incorporated into the RISC more efficiently was less tightly paired to its complement and began with an A•U pair, whereas the less effectively incorporated strand had a G•C terminus. This led to the hypothesis that the RISC preferentially accepts the strand of the siRNA that has the less stable 5′ end. To test this concept, Zamore and colleagues mutated a series of siRNAs that targeted the human Cu/Zn superoxide dismutase (SOD1) gene. By altering the stability of the base pairing at the ends (by creating a mismatch or just changing the bases), and by introducing mismatches throughout the neighbouring three nucleotides, the effective strand of the siRNA could be incorporated into the RISC. Using an independent approach, Jayasena and colleagues26 reached similar conclusions. They carried out statistical analysis on the thermodynamic profiles of numerous miRNA precursors, which revealed the presence of an unstable 5′ antisense terminus in the pre-miRNA. Notably, similar thermodynamic characteristics were present only in functional siRNAs. So, effective siRNAs duplexes can be generated by modifying the sense strand of the siRNA duplex in such a way that the antisense strand preferentially enters the RNAi pathway. In another study, Rana and colleagues28,29 emphasized two requirements for

VOLUME 5 | MAY 2004 | 3 5 7

REVIEWS

Box 2 | Considerations for generating an effective and specific siRNA Criteria

A small interfering RNA (siRNA) is defined as successful when it can provide effective and specific gene silencing (>90% reduction in protein levels) when it is used at a concentration of 1– 20 nM121. Recent studies have important implications for the design of functional siRNAs for mammalian RNAi25–27,29–31,48–50,105,122,123. On the basis of these studies, the table depicts mechanism-based rules that when applied to siRNA design are expected to show maximum suppression of target mRNA expression at the lowest possible concentration of siRNA.

Probable reason

Biophysical, thermodynamic and structural considerations Overall low to medium G+C content (30–50%)

Facilitates interaction with RISC and unwinding

Low internal stability at the 5′ antisense strand

Promotes antisense-strand selection by RISC

High internal stability at the 5′ sense strand

Blocks sense-strand selection by RISC

Absence of internal repeats or palindromes

Increases the concentration of functional, stable hairpins

A-form helix between siRNAs and target mRNA

Enhances RNA–RNA interactions and promotes cleavage

Base preferences at specific positions in the sense strand Presence of an A at position 3 and 19* of sense strand

Promotes antisense-strand selection by RISC

Absence of a G or C at position 19 of sense strand

Promotes antisense-strand selection by RISC

Presence of a U at position 10 of sense strand

Promotes RISC mediated cleavage of mRNA and dissociation of the RISC–siRNA complex

Absence of a G at position 13 of sense strand

Promotes efficient unwinding

Enhancing specificity of siRNA-mediated gene silencing Perform stringent homology searches

Minimizes potential nonspecific gene silencing

Avoid low-stringency sequence interactions between siRNA and 3′UTR

Minimizes potential nonspecific gene silencing

*The base preference for A at position 19 reflects the same bias that is observed for microRNA (miRNA) precursors. For example, most miRNAs contain a U at position 1 (corresponding to A in position 19 of the siRNA (small interfering RNA) sense strand). RISC, RNA-induced silencing complex; UTR, untranslated region.

efficient siRNA function: the 5′-phosphorylation of the antisense strand and the requirement for the antisenseRNA–target duplex to have an A-form helical structure. How do these studies influence the construction of effective siRNAs? Several important parameters in designing effective siRNAs have been revealed by all of the above studies, as well as from the systematic analysis30 of 180 siRNAs that targeted every other position 19-nt duplex region

1

5′-P

2

3

4

5

6

7

A

8

9 10 11 12 13 14 15 16 17 18 19

U

A

3′-HO

OH-3′ P-5′

High stability of the 5′ SS terminus blocks incorporation of SS into RISC. Suggestion: G or C at 5′ end of SS.

Low stability of the 5′ AS terminus promotes incorporation of AS into RISC. Suggestion: AU-richness at 5′ end of AS.

Low stability in this region promotes RISC-AS-mediated clevage of mRNA and might promote RISC-AS-complex release. Suggestion: U at position 10 of SS.

Sense strand (SS)

Antisense strand (AS)

Figure 2 | The generation of effective siRNA. A small interfering RNA (siRNA) is a 21–23nucleotide (nt) dsRNA that contains: a 19-nt duplexed region, symmetric 2–3-nt 3′ overhangs, and 5′-phosphate (P) and 3′-hydroxyl (OH) groups. The positions of each nucleotide in the 19-nt duplexed region of the sense strand are shown. On the basis of recently established design criteria, an effective siRNA has high stability at the 5′ terminus of the sense strand (blue box), lower stability at the 5′ antisense terminus (orange box) and at the cleavage site (purple box). In addition, the sequence-specific preferences at the following positions on the sense strand are important: the presence of an A at position 19, an A at position 3, a U at position 10 (BOX 2 lists other parameters). RISC, RNA-induced silencing complex.

358

| MAY 2004 | VOLUME 5

across 2 mRNA regions of 197 nt (of the firefly luciferase and human cyclophilin B genes; BOX 2, FIG. 2). Most of these parameters are aimed at promoting the assembly of the siRNA with the RISC, activation of the RISC and entry of the antisense strand into the RNAi pathway, followed by target-mRNA recognition and target-mRNA cleavage. The application of these rules significantly improved effective siRNA selection30, by increasing the probability that relatively lower concentrations of siRNA will be able to specifically silence target genes in vivo. As shRNA-based RNAi is becoming the preferred silencing method for several applications, an important question is whether similar rules for siRNA design will apply to plasmid-encoded shRNAs. In principle they should, because the crucial features that were identified as being necessary for the generation of potent siRNAs are conserved in the hairpins of naturally occurring miRNAs; a preliminary analysis31 also supports this prediction. What about the sequences outside the siRNA? One area that has received relatively less attention is the impact of local secondary structures in the mRNA, which might limit the accessibility of the siRNAs33–36 to their target. For example, striking differences in the silencing efficiency of several siRNAs that target different sites in the mRNA of the human coagulation trigger factor gene were observed37. Similarly, the ability of siRNA to block the expression of the insulin-like growth factor receptor (IGF1R) correlated with the accessibility of the target sequence within the transcript38. So, the ability to accurately predict secondary structures might be an important consideration in designing and producing effective siRNA. The application of an algorithm that can predict the secondary structure of mRNA from its primary sequence39 will further aid in this process.

www.nature.com/reviews/genetics

REVIEWS Screening approach. Although improved rational designs have enhanced the ability to generate effective siRNAs, they still do not ensure that a single predicted siRNA will silence the target gene with the highest accuracy. In light of these observations, efforts are underway to screen for effective siRNA sequences from a panel of available siRNAs. A ribonuclease H (RNAse H)-susceptibility method, which was originally used to identify ribozymeaccessible sites, was adapted for this purpose33,35. In this assay, an effective siRNA-oligonucleotide–mRNA duplex substrate becomes susceptible to RNAse H activity; the degree of RNAse H sensitivity of a given siRNA reflects the accessibility of the chosen site in the mRNA, and could predict how well a siRNA will perform. However, this process is generally time-consuming and tedious, and ultimately it still requires a duplex siRNA to be synthesized for experimental testing. More recently, a quantifiable procedure that uses conventional transfection experiments to rapidly identify effective siRNAs has been developed32. In this method, effective siRNAs are identified on the basis of their ability to reduce the expression of cognate sequences in an ectopically expressed target mRNA that is fused to a reporter gene. The siRNA probes that were identified in this way suppressed the expression of several genes with diverse biological functions — either in their endogenous setting or when ectopically expressed. The approach allowed not only the identification of the most effective siRNAs, but also of those that produced the partial suppression of target-gene expression. Such siRNAs would be useful when varying degrees of gene silencing might result in unique phenotypes. For example, shRNAs showing varying levels of TP53 gene suppression generated distinct tumour phenotypes in vivo40. These siRNAs would also be useful when lethality that is associated with the complete suppression of critical genes is of concern. However, in each case the specificity of gene silencing will need to be monitored. The development of a microarray-based cell-transfection (‘RNAi microarrays’) method has recently made it possible to automate the screening for effective siRNA probes (FIG. 3). By using this approach, loss-of-function phenotypes that are generated by RNAi in mammalian cells can potentially be scored at the single-cell level by automated microscopes with digital imaging analysis (for recent reviews, see REFS 41,42).

FEATURE

A spot on a microarray that has a defined shape, size and intensity.

NATURE REVIEWS | GENETICS

Determining siRNA specificity with DNA microarrays. The observation that even a single base mismatch between an siRNA and its mRNA target abolishes gene silencing43,44 indicates that siRNAs are highly specific. It is this property that allows allele-specific gene silencing45,46. However, siRNA-induced gene silencing has been examined only for a small number of genes, and it is therefore vital to determine that the observed phenotypes reflect the specific knockdown of the target gene. An siRNA can have several types of nonspecific effects, including degradation of partially complementary mRNA due to crosshybridization, translational silencing through an miRNA effect23 and, in some cases, the induction of genes that result in a nonspecific interferon response12, 13.

a

EGFP Target

+

RFP

+ RNAi probes: plasmid-encoded shRNA or synthetic siRNA Transfer mixture to individual wells

b

Spotting on glass slide

RNAi microarray

Reverse transfection

c

Image acquisition EGFP

RFP

MERGE

Image analysis EGFP

RFP MERGE

EGFP-specific RNA nonspecific RNA

Figure 3 | Microarray-based screening for effective siRNA. A schematic that shows the manufacture and analysis of an ‘RNAi microarray’. a | A mixture is prepared that contains: plasmid-based short hairpin RNA (shRNA) or synthetic small interfering RNA (siRNA), and constructs that express the cognate target gene fused to enhanced green fluorescent protein (EGFP). The red fluorescent protein (RFP) is used as an internal control. b | The mixtures are robotically arrayed on glass slides. The slides are then overlaid with a monolayer of mammalian cells; only the cells growing in close proximity to the DNA spots are transfected125, which results in spatially distinct groups of transfected cells within a lawn of untransfected cells. c | The effects of gene silencing are quantified by using a laser scanner to monitor the reduction in EGFP intensity. A portion of the image of the EGFP- and RFP-expressing microarray is shown. Each FEATURE is 500 µM in diameter and is separated from its neighbour by 750 µM. The right-most panel is a magnified view of a feature. In the bottom panel of part c is an RNAi microarray showing that an effective EGFP-specific siRNA results in the suppression of EGFP but not RFP expression. A nonspecific siRNA did not affect either EGFP or RFP expression. The RNAi microarray platform is flexible, robust, miniaturized and cost-effective, and it should be possible to array and analyse as many as 10,000 individual siRNA spots per glass slide. In a typical laboratory setting conventional transfections in multi-well plates are sufficient for screening effective siRNA for a smaller number of genes.

VOLUME 5 | MAY 2004 | 3 5 9

REVIEWS These concerns require that the extent of siRNA specificity be explored on a genome-wide scale. Three independent studies used the unbiased, DNA-microarray approach to measure changes at the mRNA level that were the result of siRNA-mediated gene silencing47–49. If the siRNA only produced specific effects, then it was predicted that different siRNAs against the same gene would generate similar gene-expression signatures, and siRNAs designed against different genes would have little overlap between their signatures. Indeed, this was shown to be the case by Brown and colleagues47: a specific siRNA silenced endogenous GFP expression by >70%, but did not affect the expression levels of ~36,000 human genes more than twofold. In related studies, Semizarov and colleagues48 confirmed the specificity of siRNAs by showing that the molecular signatures for siRNAs targeted against different regions of the same endogenous human genes (for example RB, AKT1 and PLK1) were comparable, but that the signatures of different genes were not. However, using similar approaches Linsley and colleagues49 showed that the siRNAs were not always target-specific: they observed the direct silencing of non-targeted genes with limited sequence similarity to the siRNA, in addition to siRNAspecific effects. In some cases the nonspecific gene silencing is assumed to have been caused by crosshybridization to transcripts that were identical in a stretch of 11–15 contiguous nucleotides49. So why did the first two studies fail to observe nonspecific effects? The are two possibile reasons. First, the siRNA-mediated nonspecific effects could have been masked by experimental noise that is commonly associated with microarray experiments and, second, nonspecific effects might only occur only for poorly (albeit inadvertently) designed siRNAs. Recently, sequence-dependent nonspecific effects of siRNA have also been observed at the protein level50, which underscores the importance of not only improving siRNA designs but also for carrying out stringent homology searches against the genome databases (BOX 2). Although it might not be generally feasible to use microarrays to assess the specificity of every siRNA, it would be wise to confirm any phenotype that is caused by an siRNA by using a second, independent siRNA that is directed against the same target, and/or by rescuing the loss-of-function phenotype by expressing a modified version of the target gene that has silent mutations in the siRNA-target region. Using many siRNA duplexes to silence a target gene of interest will also increase the confidence with which an observed phenotype and expression pattern can be linked to target-gene silencing. In summary, we are already beginning to see overwhelming progress in the development and application of algorithms for designing effective siRNAs (BOX 2 and references therein). These algorithms will be gradually refined as more information is gathered on RNAi biochemistry, the impact of the secondary structure of target mRNAs on siRNA accessibility, and the nonspecific silencing of unintended transcripts. Until then the most powerful and specific siRNAs are likely to be identified by combining rational designs with appropriate screening methods32.

360

| MAY 2004 | VOLUME 5

Delivery of siRNA

Targeting cells in culture. The efficient delivery of siRNAs into mammalian cells is a vital step in most RNAi-based gene-silencing experiments, and has undergone considerable improvements. Synthetic siRNAs can be delivered to mammalian cells in culture by electroporation or by using lipophilic agents25,51,52, and have been successfully used to silence target genes (for a list of targeted genes, see REF. 5). However, these approaches are limited by the transient nature of the response and in some cases by lipid-mediated toxicity; in this respect, efforts are underway to modify the phosphate backbone of the siRNA to minimize its charge and therefore to facilitate its entry into the cell. The use of plasmids that harbour siRNAexpression cassettes overcomes some of these limitations: these plasmids are cheap to produce and provide a continuous expression of hairpin RNAs. They are therefore very useful for analysing loss-of-function phenotypes that develop over extended periods of time. The shRNA is predicted to contain a perfectly double-stranded stem of 19–29 bp that is identical in sequence to the target mRNA; the 2 strands of the stem are connected by a loop of 6–9 bases (FIG. 1), which is removed in vivo by Dicer to generate effective siRNAs. The constitutive expression of plasmid-based shRNAs by RNA polymerase III (pol III) U6 and H1 snRNA promoters8,9,11,34,53–55, tRNA promoters56, and RNA-pol-II-based CMV (cytomegalovirus) promoters57,58 has been used successfully to obtain stable and efficient suppression of target genes5. The pol III promoter is widely used for directing the expression of shRNAs, because it is active in all cell types and efficiently directs the synthesis of small, non-coding transcripts (which are structurally close to shRNAs) that bear welldefined ends. By contrast, pol II promoters lack these properties and have therefore found limited use in mediating efficient gene silencing. Although the use of plasmid-based shRNA expression has quickly become the preferred method for targeted gene silencing, there are potential limitations. For example, the delivery of plasmid-expressed shRNA cells that are difficult to transfect, such as freshly isolated primary cells, neural cells, and stem cells, has been hampered by inefficient transfection protocols; as a result, cell lines that stably express shRNA cannot be established. The delivery of siRNA by viral transduction is one way to overcome these shortcomings. The efficient introduction and stable integration of these shRNA-expression cassettes into the host genome can be efficiently achieved by using the Moloney murine leukaemia virus40,55,59–61, the murine stem-cell virus40,62 or lentiviral63–69 vector systems. By contrast, vector systems that are based on adenoviruses do not integrate into the host genome57,70,71 and are therefore more suitable for RNAi-mediated human gene therapy. The various viral delivery systems ensure broad tropism and guarantee that shRNA-expression cassettes can be delivered into both transformed and primary mammalian cells. Lentivirus-mediated delivery has added advantages over the others; this virus can efficiently integrate into the genome of non-dividing cells, such as stem cells or terminally differentiated cells, which are refractory to

www.nature.com/reviews/genetics

REVIEWS conventional retroviral infection. In addition, transgenes carried by lentiviral vectors are resistant to silencing72. Indeed lentivirus-delivered siRNA has effectively suppressed GFP64, the pro-apoptotic BIM1 (Bcl2-interacting mediator of cell death) gene and the gene encoding the IL2 receptor (CD25)63. Inhibition of HIV-1 infection in human T cells by the lentivirus-mediated delivery of siRNA against CCR5 (the chemokine (C-C motif) receptor 5)69 has generated tremendous excitement owing to the potential therapeutic application of virus-mediated RNAi delivery. Recently, however, retrovirus-mediated gene therapy for treatment for X-linked severe combined immunodeficiency (X-SCID) resulted in the development of T-cell leukaemia in two children73–75 owing to the insertional mutagenesis of the vector in the proto-oncogene LMO2 (LIM-domain-only 2) gene. These studies reinforce the need to test lentiviral delivery systems that are associated with infrequent insertional mutagenesis, and to perhaps explore the use of adenoviral systems that do not integrate into the host genome. Nevertheless, the suppression of cognate target genes by virus-mediated RNAi delivery has widespread use, which indicates the potential therapeutic applications of RNAi-based methods. Delivery to live animals. The use of RNAi in living mice has been widely reported. The germline transmission of cells — embryonic stem (ES) cells in mice76 and fertilized eggs in rats77 — that carry an shRNA transgene has been accomplished, and in some cases mice are phenotypically identical to those that carry a null mutation in the target

gene78. In addition, the in vivo transfection of siRNA directly into the organs of postnatal mice inhibits the expression of co-transfected reporter plasmids79,80, of the endogenous pro-apoptoic Fas receptor81, and of the TH (tyrosine hydroxylase) gene in the brain82. These data indicate that, in the future, RNAi could be used to generate knockout mice more rapidly than is possible using standard methodologies. Because lentivirusmediated transduction allows mice carrying shRNA constructs to be produced63,65,76, in principle it is possible to produce efficiently an array of mutant mice with shRNA constructs that target different genes. Once these mice are systematically screened on the basis of their phenotype, the gene that is responsible for a certain phenotype can be readily identified, unlike the more laborious gene mapping that follows large-scale ENU (ethylnitrosurea) mutagenesis projects in mice83. The demonstration that RNAi can be used for targeting genes in the rat is particularly significant, as standard gene targeting is impossible in this organism owing to the lack of suitable rat ES cell lines. In future, RNAi-generated knockout rats should be useful models for unraveling the genetic roots of many human diseases, including diabetes, hypertension and neurological disorders. Although RNAi has been shown to downregulate the expression of various target genes effectively in animal models, it is still uncertain whether RNAi will be translated into effective therapy. More progress is required to improve RNAi-delivery systems and to evaluate the toxicity of exogenous RNA and its effects on endogenous RNAi processes.

Table 1| The conditional expression of shRNA in mammalian cells System

Promoter used

Description

Genes suppressed

Plasmid based; tetracycline-inducible

Pol III (7SK, U6, H1)

The tetO sequence is located immediately downstream of the TATA box. In the absence of Dox, tTR binds to tetO and represses shRNA transcription. Adding Dox results in the dissociation of tTR, which leads to the derepression/activation of the transcriptional unit and to shRNA expression.

Human β-catenin; two catalytic subunits of the PI3K, p110α and p110β

References

Plasmid based; tetracycline-inducible

Pol III (U6)

Multiple tetO sequences are incorporated between the pol III promoter elements DSE and PSE and the TATA box. This design enhances the cooperative binding of tTR to the tetO sequences and simultaneously blocks the communication between each of these elements. Dox results in dissociation of tTR, which leads to the derepression/activation of the transcriptional unit and to shRNA expression.

CXC chemokine receptor-4 (CXCR4)

89

Lentivirus-based, tetracycline-inducible

Pol III (H1)

Uses a tetracycline-controlled hybrid protein, tTR–KRAB. In the absence of Dox, tTR-KRAB binds to tetO and suppresses shRNA transcription. Adding Dox results in dissociation of tTR–KRAB, which leads to derepression/activation of the transcriptional unit and to shRNA expression.

human TP53, Lamin A/C

87

Moloney murine leukaemia virus (MoMLV)-based, ecdysone-inducible

Pol III (U6)

Comprises three retroviral vectors: two expressing transcription factors VgEcR and RXR, and one expressing a Gal4-chimeric transactivator fusion gene under the control of an inducible promoter. On induction by ecdysone the transcription factors dimerize and bind to the inducible promoter to activate the GAL4-transactivator; in turn this binds to the four GAL4 DNA-binding sites and activates the U6 promoter, which drives shRNA expression.

Mouse MyoD; human TP53

92

85,86

DSE, distal sequence element; Dox, Doxycyclin; GAL4 = N-terminal fragment of a yeast transcriptional activator; KRAB = Kruppel-associated-box family of transcriptional repressors; PI3K, phosphatidylinositol 3-kinase; PSE, proximal sequence element; Pol III, RNA polymerase III; RXR = Retinoid X receptor; shRNA, short hairpin RNA; TATA, TATA box; tetO, tetracycline operator; tTR, the tetracyclin repressor protein; VgEcR = modified ecdysone receptor.

NATURE REVIEWS | GENETICS

VOLUME 5 | MAY 2004 | 3 6 1

REVIEWS

Library of pooled viral vectors (barcoded)

Array-formatted RNAi libraries

Synthetic siRNAs

Plasmid-based shRNAs

Viral vectors encoding shRNAs Individuals or pools

Transfect mammalian cells

Generate virus

Transduce mammalian cells at low MOIs Selection Phenotype analysis PCR-amplify barcodes Label with fluorescent dyes If pooled recover and identify shRNA insert

Hybridize to microarray containing barcode oligonucleotides

Identify and validate responsible siRNA

Figure 4 | Large-scale RNAi screens in cells. a | Array-formatted RNA interference (RNAi) libraries consist of individual RNAi probes that target different genes and that are placed at defined locations in a multi-well dish. Libraries can comprise synthetic small interfering RNAs (siRNAs), plasmid-based short hairpin RNAs (shRNAs) or viral vector-encoded shRNAs. The non-viral vectors are directly transfected into mammalian cells, either individually in each well of a multi-well dish, or as pools of reasonable size. By contrast, viral vectors are generated from pooled vector libraries and can be used for infecting any cell line. Common phenotypes that are scored include changes in general cell morphology, cell proliferation or death; more sophisticated phenotype analyses require the use of fluorescent or luminescent reporters. Cells that show phenotype changes can be screened if the gene-silencing vector is marked, for example, with a molecular barcode. The relative abundance of each hairpin-containing vector in any cell population can be quantified by labeling the PCR product with fluorescent dyes (such as Cy5 or Cy3); the PCR products are then hybridized to microarrays containing barcode-complementary oligonucleotides and their relative abundance is compared to that detected in control cells that have been exposed to the same shRNA library, but not to the biological stimulus of interest (for example, drug treatment or genetic mutations). MOI, multiplicity of infection.

ORTHOTOPIC

The transplantation of foreign tumour cells into another species at its normal anatomical position.

362

| MAY 2004 | VOLUME 5

transferase-1) gene in human cancer cells resulted in growth arrest88, the silencing of the endogenous CXCR4 (CXC chemokine receptor-4) in breast cancer cells resulted in a significant inhibition of breast-cancer-cell migration in vitro89, and knockdown of the β-catenin gene in colorectal cancer cells resulted in cell-cycle and growth arrest86. Specific and inducible downregulation of human TP53, LMNA (Lamin A/C)87, and of the genes encoding two catalytic subunits of the human phosphatidylinositol 3-kinase (PI3K), p110α and p110β, in prostate cancer cells and in an ORTHOTOPIC prostate tumour metastasis model has also been accomplished85. Conditional shRNA expression was stable and maintained for about 56 days in vivo85. A limitation of the tetracycline-inducible system is a relatively high background of expression in the uninduced state in certain cell lines90,91, and toxicity problems in some cases limit studies involving both cultured cells and animals. However, taking these limitations into consideration, a system that allows ecdysone-inducible synthesis of shRNAs under the control of a modified pol-III-specific U6 promoter has recently been developed92. This system is tightly regulated in both mouse and human cells with undetectable background expression levels in the absence of the inducer, which is consistent with previous observations93. The specific downregulation of human TP53, mouse MyoD and helix-loop-helix Id genes (V.M. and S. Gupta, unpublished data) followed the dosage- and time-dependent kinetics of induction with undetectable background suppression in the absence of the inducer. The inducibility of this system can be reversed by withdrawing the inducer, as observed by the reappearance of protein and a restoration of the original cell phenotype. The generation of an inducible pol III promoter opens up the possibility of tissue- or cell-specific regulation of shRNAs by expressing the GAL4-activator under the control of defined tissue-specific promoters. This is likely to have widespread applications in both cultured cells and in living animals.

Expression of siRNA

Large-scale RNAi screens in cells

Inducible RNAi. The existing method of gene suppression by the constitutive expression of shRNAs allows the consequences of stably silencing genes to be analysed, but significantly limits the study of genes that are essential for cell survival, cell-cycle regulation and development. In addition, the gross suppression of a gene over longer periods can result in compensatory or even nonphysiological responses that mask the true biological consequence of a functional knockdown. Generating the inducible regulation of RNAi should be able to circumvent this problem. Pol II promoters have been used to inducibly express long hairpin RNAs that target the β-galactosidase gene in Drosophila melanogaster84, but the development of inducible RNAi for mammalian cells is fairly recent (TABLE 1). Several groups have developed inducible siRNAs, the expression of which is controlled by the tetracycline- or doxycycline-regulated form of the pol III (U6 or H1) promoter85–89. For example, the controlled knockdown of the DNMT1 (DNA methyl-

RNAi has emerged as a powerful approach for studying loss-of-function phenotypes on a genome-wide level. Initially, rapid progress in genome-wide RNAi screens was made in model organisms such as worms and flies primarily owing to the ease with which dsRNA could be delivered: dsRNA could be administered to worms by feeding them bacteria that express dsRNA, soaking them in a dsRNA solution, or by the direct microinjection of in-vitro-transcribed long dsRNA. These advantages are coupled to the absence of the undesirable effects that are caused by long dsRNA in mammals. In C. elegans, RNAi screens have allowed the identification of genes that affect worm development94,95, mitochondrial function96, regulation of fat storage97 and suppression of mutagenesis98. Such approaches have also been carried out in D. melanogaster cell lines99–101, in which it has been particularly successful in identifying components of the Hedgehog (Hh) signaling pathway99. To uncover new Hh-pathway regulators, Lum and colleagues generated a

www.nature.com/reviews/genetics

REVIEWS

TRAIL-INDUCED APOPTOSIS

The tumour necrosis factor (TNF)-receptor-related apoptosis-inducing ligand (TRAIL) is a secreted protein ligand that binds to its receptor and induces apoptosis selectively in tumour cells but not in normal cells. SYNTHETIC LETHAL MUTATIONS

Two mutations are considered to be synthetically lethal if in combination they result in cell death, whereas either alone leads to a viable cell. BARCODE

A short stretch of DNA that functions as a unique molecular fingerprint for each short hairpin RNA in a library. The molecular barcode can be identified and recovered by PCR amplification, by using primers that are specific to each barcode.

library of dsRNAs that represented the entire set of D. melanogaster kinases and phosphatases, as well as approximately 43% of all predicted D. melanogaster genes, and carried out rapid primary screens with small pools of dsRNAs. The in vivo functions of selected candidates were validated by phenotypic analysis of embryos into which corresponding dsRNAs were injected. These screens identified four new genes that are involved in Hh signaling. More recently, the availability of well-annotated human and mouse genomic sequences102–104, and advances in the efficient design and delivery of siRNA to mammalian cells, has opened up the possibility of applying RNAi to genome-wide loss-of-function genetic screens in mammalian systems. The first demonstration of a successful RNAi-based screening strategy in mammalian cells came from the studies of Aza-Blanc and colleagues105. They used a chemically synthesized siRNA library to individually target 510 human genes (which corresponds to 380 predicted kinases and 130 other predicted proteins) to identify modulators of TRAIL-INDUCED APOPTOSIS, using cell viability as the phenotypic assay. The authors identified both known and new molecules that regulate TRAIL-induced apoptosis, as well as new connections between components of known molecular pathways and apoptosis. The use of RNAi libraries in mammalian systems has been largely limited to specific protein families, because of technical and practical issues that are associated with generating large synthetic siRNA and/or vector-based shRNA expression libraries. However, efforts are underway to develop larger siRNA-expression libraries106–109.

Box 3 | Examples of cell-based RNAi screens in mammalian systems By using plasmid-based short hairpin RNA (shRNA) vectors to target 50 human de-ubiquitylating enzymes, Bernard and colleagues110 identified new regulators of nuclear factor κB (NF-κB) that are deregulated in cancer. They identified the familial CYLD (cylindromatosis) tumour suppressor gene, previously of unknown function, as a new negative regulator of NF-κB that leads to increased resistance to apoptosis. Zheng and colleagues106 have developed a single-step PCR approach to generate plasmid-based shRNA expression libraries in a high-throughput manner. By using two small interfering RNA (siRNA) sequences per gene they targeted >8,000 genes and identified known components of the NF-κB signaling pathway and genes that previously had unrecognized roles in regulating transcriptional activity of NF-κB. Bernard and colleagues109 generated and applied array-formatted libraries of retroviral vectors that encoded sequence-verified shRNAs against 7,914 individual human genes. The library was organized into 83 pools. The use of high-titre virus from each pool identified one known and five new modulators of p53-dependent proliferation arrest in human cells. The main advantage of using an arrayed library is the ability to assign a gene to a particular phenotype that is detected in a screen. A drawback of this approach is that it is tedious and expensive because hairpin vectors have to be recovered from numerous clones and retested individually. To circumvent this problem, Bernard and colleagues used an alternative strategy that is known as the ‘siRNA barcode screen’: individual clones in the library contain a unique tag (barcode), which can be recovered by PCR amplification and identified by quantifiable hybridization to a DNA microarray that consists of barcode-complementary DNA fragments. Such a barcoding method has been very successful in yeast genetic screens124 . Hannon and colleagues127 describe the construction and application of a retroviralbased barcoded shRNA expression library that targets almost 10,000 human and more than 5,000 murine genes. Bacterial mating allows the shRNA-encoding cassettes to be shuttled into various customized vectors.

NATURE REVIEWS | GENETICS

The development and use of these libraries in cellbased RNAi screens are already generating important results106,109,110,127 (FIG. 4, BOX 3). Although RNAi is a powerful new tool for carrying out genome-wide loss-of–function analysis, there are also concerns and limitations in the use of this technology. In some cases, the insufficient silencing of a gene might not generate an observable phenotype, thereby contributing to the false-negative rate of the screen. Indeed, a recent study in C. elegans showed false-negative rates of 10–30% between two repeat screens111. In addition, there can be considerable experimental variability in the degree of knockdown that is achieved using RNAi. This experimental noise can make more subtle effects difficult to discern, and can make detailed and precise genetic analysis difficult. It should, however, be realized that large-scale analyses are still at an incipient stage and parallel developments in generating more effective siRNA designs, automated phenotype analysis and bioinformatics tools will bolster the feasibility of efficiently using RNAi to analyse mammalian genomes. Future prospects and conclusions

Many of the important obstacles that have plagued the RNAi field in the past have now been alleviated. This is due to recent developments in several areas: the ability to predict potent siRNAs against any gene, the large-scale identification of effective siRNAs, a wider choice of RNAi delivery systems both in vitro and in vivo, and the inducible suppression of endogenous genes. The ability to design effective siRNAs against any gene has already jump-started genome-wide functional RNAi screens in mammalian systems, which had lagged behind invertebrate models. However, at present, the lack of appropriate RNAi libraries is the main restriction on conducting such large-scale loss-of-function genetic screens. Because the production of such a large-scale resource is beyond most individual laboratories, one solution would be to have the requisite reagents provided by academic core facilities or commercial entities. This approach has already proved to be successful for other genomic technologies, such as DNA-microarray-based gene-expression profiling, so its application to RNAi experiments is likely to be enthusiastically embraced by the research community. With the eventual creation of appropriate siRNA libraries, it should be possible to undertake informative RNAi screens in mammalian cells, including the identification of ‘SYNTHETIC LETHAL’ interactions that should aid in the discovery of new cancer drug targets. For example, particular genes that when silenced show synthetic lethal interactions with cancer-specific mutations in a second gene would make ideal drug targets. In large-scale RNAi screens, for which pooling strategies are preferred, cell death would result in the loss of the vector expressing the siRNA; in such cases the use of BARCODED siRNA libraries109,127 would be particularly useful. In living mice, the success of lentivirus-mediated gene-specific knockdowns and of germline transmission of RNAi63,65,76 should make it feasible to create an array of mutant mice with shRNA constructs that target different genes, an approach that would be more efficient than

VOLUME 5 | MAY 2004 | 3 6 3

REVIEWS traditional homologous recombination. This mutant library would be a particularly powerful resource if gene knockdown in these animals could be controlled in a temporal- and tissue-specific fashion by using an inducible system — it would then match existing conditional knockout strategies while being more rapid, costeffective and less labour-intensive. Improving the efficiency of RNAi in mammals is likely to influence the study of gene pathways. In combination with genome-wide expression profiling, RNAi provides an effective means to both identify and validate the components of a signaling pathway that are associated with the silenced gene. For example, silencing of protooncogene FRA1 revealed the induction of the CD44 and cMET genes, as assayed by using microarrays112. DNA microarrays have also been used to identify targets of the

1. 2.

3. 4. 5.

6.

7.

8.

9.

10.

11.

12.

13.

14.

15.

16. 17. 18.

364

Sharp, P. A. RNAi and double-strand RNA. Genes Dev. 13, 139–141 (1999). Hammond, S. M., Caudy, A. A. & Hannon, G. J. Posttranscriptional gene silencing by double-stranded RNA. Nature Rev. Genet. 2, 110–119 (2001). Hannon, G. J. RNA interference. Nature 418, 244–251 (2002). McManus, M. T. & Sharp, P. A. Gene silencing in mammals by small interfering RNAs. Nature Rev. Genet. 3, 737–747 (2002). Dykxhoorn, D. M., Novina, C. D. & Sharp, P. A. Killing the messenger: short RNAs that silence gene expression. Nature Rev. Mol. Cell Biol. 4, 457–467 (2003). Elbashir, S. M., Lendeckel, W. & Tuschl, T. RNA interference is mediated by 21- and 22-nucleotide RNAs. Genes Dev. 15, 188–200 (2001). Shows for the first time that chemically synthesized short dsRNA molecules of 21-22 nt — known as small interfering RNAs (siRNAs) — could be used to target mammalian genes by RNAi while evading the interferon responses. This was confirmed independently by reference 7. Caplen, N. J., Parrish, S., Imani, F., Fire, A. & Morgan, R. A. Specific inhibition of gene expression by small doublestranded RNAs in invertebrate and vertebrate systems. Proc. Natl Acad. Sci. USA. 98, 9742–9747 (2001). Paddison, P. J., Caudy, A. A., Bernstein, E., Hannon, G. J. & Conklin, D. S. Short hairpin RNAs (shRNAs) induce sequence-specific silencing in mammalian cells. Genes Dev. 16, 948–958 (2002). This work, together with references 9–11 and 53–55, shows that plasmids are cost-effective; they also provide a continuous expression of hairpin RNAs that is useful for analysing loss-of-function phenotypes that develop over extended periods of time. Sui, G. et al. A DNA vector-based RNAi technology to suppress gene expression in mammalian cells. Proc. Natl Acad. Sci. USA 99, 5515–5520 (2002). Brummelkamp, T. R. & Bernards, R. New tools for functional mammalian cancer genetics. Nature Rev. Cancer 3, 781–789 (2003). Paul, C. P., Good, P. D., Winer, I. & Engelke, D. R. Effective expression of small interfering RNA in human cells. Nature Biotechnol. 20, 505–508 (2002). Bridge, A. J., Pebernard, S., Ducraux, A., Nicoulaz, A. L. & Iggo, R. Induction of an interferon response by RNAi vectors in mammalian cells. Nature Genet. 34, 263–264 (2003). Sledz, C. A., Holko, M., de Veer, M. J., Silverman, R. H. & Williams, B. R. Activation of the interferon system by shortinterfering RNAs. Nature Cell Biol. 5, 834–839 (2003). Shuey, D. J., McCallus, D. E. & Giordano, T. RNAi: genesilencing in therapeutic intervention. Drug Discov. Today 7, 1040–1046 (2002). Silva, J. M., Hammond, S. M. & Hannon, G. J. RNA interference: a promising approach to antiviral therapy? Trends Mol. Med. 8, 505–508 (2002). Kim, V. N. RNA interference in functional genomics and medicine. J. Korean Med. Sci. 18, 309–318 (2003). Capecchi, M. R. The new mouse genetics: altering the genome by gene targeting. Trends Genet. 5, 70–76 (1989). Scherer, L. J. & Rossi, J. J. Approaches for the sequencespecific knockdown of mRNA. Nature Biotechnol. 21, 1457–1465 (2003). A review that highlights the use of RNA-based methods such as antisense vectors, catalytic RNA molecules (ribozymes) and catalytic DNA molecules (DNAzymes) to suppress gene expression.

| MAY 2004 | VOLUME 5

mouse helix-loop-helix protein Id, which is involved in supporting tumour angiogenesis113, and this has shown that siRNA-mediated loss of Id gene function phenocopied the angiogenic defect that is observed in Id knockout mice (V.M., unpublished data). In the future, we should be able to dissect specific pathways by treating cells sequentially with siRNAs that are targeted against various genes in the pathway and then assaying which genes are affected by each knockdown. An advantage of DNAmicroarray analyses is that, unlike loss-of-function screens, they do not completely rely on genes that do not have gross morphological phenotypes. An important direction for future research will be to integrate the data that are emerging from large-scale RNAi screens with that obtained from other approaches, such as gene-expression profiling and protein-interaction mapping.

19. Opalinska, J. B. & Gewirtz, A. M. Nucleic-acid therapeutics: basic principles and recent applications. Nature Rev. Drug Discov. 1, 503–514 (2002). 20. Bernstein, E., Caudy, A. A., Hammond, S. M. & Hannon, G. J. Role for a bidentate ribonuclease in the initiation step of RNA interference. Nature 409, 363–366 (2001). Describes the identification of the ribonuclease III activity of the evolutionary conserved Dicer enzyme, which is required for processing long dsRNA into 21–22nt fragments (siRNA). 21. Nykanen, A., Haley, B. & Zamore, P. D. ATP requirements and small interfering RNA structure in the RNA interference pathway. Cell 107, 309–321 (2001). 22. Martinez, J., Patkaniowska, A., Urlaub, H., Luhrmann, R. & Tuschl, T. Single-stranded antisense siRNAs guide target RNA cleavage in RNAi. Cell 110, 563–574 (2002). 23. Doench, J. G., Petersen, C. P. & Sharp, P. A. siRNAs can function as miRNAs. Genes Dev. 17, 438–442 (2003). 24. Zeng, Y., Yi, R. & Cullen, B. R. MicroRNAs and small interfering RNAs can inhibit mRNA expression by similar mechanisms. Proc. Natl Acad. Sci. USA 100, 9779–9784 (2003). 25. Elbashir, S. M., Harborth, J., Weber, K. & Tuschl, T. Analysis of gene function in somatic mammalian cells using small interfering RNAs. Methods 26, 199–213 (2002). 26. Khvorova, A., Reynolds, A. & Jayasena, S. D. Functional siRNAs and miRNAs exhibit strand bias. Cell 115, 209–216 (2003). 27. Schwarz, D. S. et al. Asymmetry in the assembly of the RNAi enzyme complex. Cell 115, 199–208 (2003). References 26 and 27 describe, for the first time, the mechanism of bias in siRNA strand selection that occurs during siRNA–RISC assembly and activation, and that influences the overall efficiency of RNAi. This information has contributed important parameters to the rational design of effective siRNAs. 28. Chiu, Y. L. & Rana, T. M. RNAi in human cells: basic structural and functional features of small interfering RNA. Mol. Cell 10, 549–561 (2002). 29. Chiu, Y. L. & Rana, T. M. siRNA function in RNAi: a chemical modification analysis. RNA 9, 1034–1048 (2003). 30. Reynolds, A. et al. Rational siRNA design for RNA interference. Nature Biotechnol. 22, 326–330 (2004). The authors identified important features that are required to design effective siRNAs by carrying out a systematic analysis of 180 siRNAs that target the mRNAs of two genes. 31. Ui-Tei, K. et al. Guidelines for the selection of highly effective siRNA sequences for mammalian and chick RNA interference. Nucleic Acids Res. 32, 936–948 (2004). 32. Kumar, R., Conklin, D. S. & Mittal, V. High-throughput selection of effective RNAi probes for gene silencing. Genome Res. 13, 2333–2340 (2003). 33. Lee, N. S. et al. Expression of small interfering RNAs targeted against HIV-1 rev transcripts in human cells. Nature Biotechnol. 20, 500–505 (2002). 34. Yu, J. Y., Taylor, J., DeRuiter, S. L., Vojtek, A. B. & Turner, D. L. Simultaneous inhibition of GSK3α and GSK3β using hairpin siRNA expression vectors. Mol. Ther. 7, 228–236 (2003). 35. Vickers, T. A. et al. Efficient reduction of target RNAs by small interfering RNA and RNase H-dependent antisense agents. A comparative analysis. J. Biol. Chem. 278, 7108–7118 (2003). 36. Kretschmer-Kazemi Far, R. & Sczakiel, G. The activity of siRNA in mammalian cells is related to structural target accessibility: a comparison with antisense oligonucleotides. Nucleic Acids Res. 31, 4417–4424 (2003).

37. Holen, T., Amarzguioui, M., Wiiger, M. T., Babaie, E. & Prydz, H. Positional effects of short interfering RNAs targeting the human coagulation trigger tissue factor. Nucleic Acids Res. 30, 1757–1766 (2002). 38. Bohula, E. A. et al. The efficacy of small interfering RNAs targeted to the type 1 insulin-like growth factor receptor (IGF1R) is influenced by secondary structure in the IGF1R transcript. J. Biol. Chem. 278, 15991–15997 (2003). 39. Ding, Y. & Lawrence, C. E. A statistical sampling algorithm for RNA secondary structure prediction. Nucleic Acids Res. 31, 7280–7301 (2003). 40. Hemann, M. T. et al. An epi-allelic series of p53 hypomorphs created by stable RNAi produces distinct tumor phenotypes in vivo. Nature Genet 33, 396–400 (2003). 41. Carpenter, A. E. & Sabatini, D. M. Systematic genome-wide screens of gene function. Nature Rev. Genet. 5, 11–22 (2004). 42. Mousses, S. et al. RNAi microarray analysis in cultured mammalian cells. Genome Res. 13, 2341–2347 (2003). 43. Elbashir, S. M., Martinez, J., Patkaniowska, A., Lendeckel, W. & Tuschl, T. Functional anatomy of siRNAs for mediating efficient RNAi in Drosophila melanogaster embryo lysate. EMBO J. 20, 6877–6888 (2001). 44. Tuschl, T., Zamore, P. D., Lehmann, R., Bartel, D. P. & Sharp, P. A. Targeted mRNA degradation by doublestranded RNA in vitro. Genes Dev. 13, 3191–3197 (1999). 45. Miller, V. M. et al. Allele-specific silencing of dominant disease genes. Proc. Natl Acad. Sci. USA 100, 7195–7200 (2003). 46. Gonzalez-Alegre, P., Miller, V. M., Davidson, B. L. & Paulson, H. L. Toward therapy for DYT1 dystonia: allelespecific silencing of mutant TorsinA. Ann. Neurol. 53, 781–787 (2003). 47. Chi, J. T. et al. Genomewide view of gene silencing by small interfering RNAs. Proc. Natl Acad. Sci. USA 100, 6343–6346 (2003). 48. Semizarov, D. et al. Specificity of short interfering RNA determined through gene expression signatures. Proc. Natl Acad. Sci. USA 100, 6347–6352 (2003). 49. Jackson, A. L. et al. Expression profiling reveals off-target gene regulation by RNAi. Nature Biotechnol. 21, 635–637 (2003). These authors used a DNA microarray approach to measure changes at the mRNA level that result from siRNA-mediated gene silencing. They showed, for the first time, that siRNAs can silence non-targeted genes that have limited sequence similarity to the siRNA. 50. Scacheri, P. C. et al. Short interfering RNAs can induce unexpected and divergent changes in the levels of untargeted proteins in mammalian cells. Proc. Natl Acad. Sci. USA 101, 1892–1897 (2004). 51. McManus, M. T. et al. Small interfering RNA-mediated gene silencing in T lymphocytes. J. Immunol. 169, 5754–5760 (2002). 52. Kishida, T. et al. Sequence-specific gene silencing in murine muscle induced by electroporation-mediated transfer of short interfering RNA. J. Gene Med. 6, 105–110 (2004). 53. Miyagishi, M. & Taira, K. U6 promoter-driven siRNAs with four uridine 3′ overhangs efficiently suppress targeted gene expression in mammalian cells. Nature Biotechnol. 20, 497–500 (2002). 54. McManus, M. T., Petersen, C. P., Haines, B. B., Chen, J. & Sharp, P. A. Gene silencing using micro-RNA designed hairpins. RNA 8, 842–850 (2002). 55. Brummelkamp, T. R., Bernards, R. & Agami, R. A system for stable expression of short interfering RNAs in mammalian cells. Science 296, 550–553 (2002).

www.nature.com/reviews/genetics

REVIEWS 56. Kawasaki, H. & Taira, K. Short hairpin type of dsRNAs that are controlled by tRNA(Val) promoter significantly induce RNAimediated gene silencing in the cytoplasm of human cells. Nucleic Acids Res. 31, 700–707 (2003). 57. Xia, H., Mao, Q., Paulson, H. L. & Davidson, B. L. siRNAmediated gene silencing in vitro and in vivo. Nature Biotechnol. 20, 1006–1010 (2002). 58. Shinagawa, T. & Ishii, S. Generation of Ski-knockdown mice by expressing a long double-strand RNA from an RNA polymerase II promoter. Genes Dev. 17, 1340–1345 (2003). 59. Paddison, P. J. & Hannon, G. J. RNA interference: the new somatic cell genetics? Cancer Cell 2, 17–23 (2002). 60. Barton, G. M. & Medzhitov, R. Retroviral delivery of small interfering RNA into primary cells. Proc. Natl Acad. Sci. USA 99, 14943–14945 (2002). 61. Liu, C. M., Liu de, P., Dong, W. J. & Liang, C. C. Retrovirus vector-mediated stable gene silencing in human cell. Biochem. Biophys. Res. Commun. 313, 716–720 (2004). 62. Brummelkamp, T. R., Bernards, R. & Agami, R. Stable suppression of tumorigenicity by virus-mediated RNA interference. Cancer Cell 2, 243–247 (2002). 63. Rubinson, D. A. et al. A lentivirus-based system to functionally silence genes in primary mammalian cells, stem cells and transgenic mice by RNA interference. Nature Genet. 33, 401–406 (2003). 64. Stewart, S. A. et al. Lentivirus-delivered stable gene silencing by RNAi in primary cells. RNA 9, 493–501 (2003). 65. Tiscornia, G., Singer, O., Ikawa, M. & Verma, I. M. A general method for gene knockdown in mice by using lentiviral vectors expressing small interfering RNA. Proc. Natl Acad. Sci. USA 100, 1844–1848 (2003). 66. Van den Haute, C., Eggermont, K., Nuttin, B., Debyser, Z. & Baekelandt, V. Lentiviral vector-mediated delivery of short hairpin RNA results in persistent knockdown of gene expression in mouse brain. Hum. Gene Ther. 14, 1799–1807 (2003). 67. Scherr, M., Battmer, K., Ganser, A. & Eder, M. Modulation of gene expression by lentiviral-mediated delivery of small interfering RNA. Cell Cycle 2, 251–257 (2003). 68. Abbas-Terki, T., Blanco-Bose, W., Deglon, N., Pralong, W. & Aebischer, P. Lentiviral-mediated RNA interference. Hum. Gene Ther. 13, 2197–2201 (2002). 69. Qin, X. F., An, D. S., Chen, I. S. & Baltimore, D. Inhibiting HIV-1 infection in human T cells by lentiviral-mediated delivery of small interfering RNA against CCR5. Proc. Natl Acad. Sci. USA 100, 183–188 (2003). 70. Shen, C., Buck, A. K., Liu, X., Winkler, M. & Reske, S. N. Gene silencing by adenovirus-delivered siRNA. FEBS Lett. 539, 111–114 (2003). 71. Arts, G. J. et al. Adenoviral vectors expressing siRNAs for discovery and validation of gene function. Genome Res. 13, 2325–2332 (2003). 72. Pfeifer, A., Ikawa, M., Dayn, Y. & Verma, I. M. Transgenesis by lentiviral vectors: lack of gene silencing in mammalian embryonic stem cells and preimplantation embryos. Proc. Natl Acad. Sci. USA 99, 2140–2145 (2002). 73. Williams, D. A. & Baum, C. Medicine. Gene therapy—new challenges ahead. Science 302, 400–401 (2003). 74. Hacein-Bey-Abina, S. et al. LMO2-associated clonal T cell proliferation in two patients after gene therapy for SCID-X1. Science 302, 415–419 (2003). 75. Marshall, E. Gene therapy. Second child in French trial is found to have leukemia. Science 299, 320 (2003). 76. Carmell, M. A., Zhang, L., Conklin, D. S., Hannon, G. J. & Rosenquist, T. A. Germline transmission of RNAi in mice. Nature Struct. Biol. 10, 91–92 (2003). 77. Hasuwa, H., Kaseda, K., Einarsdottir, T. & Okabe, M. Small interfering RNA and gene silencing in transgenic mice and rats. FEBS Lett. 532, 227–230 (2002). 78. Kunath, T. et al. Transgenic RNA interference in ES cellderived embryos recapitulates a genetic null phenotype. Nature Biotechnol. 21, 559–561 (2003). 79. McCaffrey, A. P. et al. RNA interference in adult mice. Nature 418, 38–39 (2002). 80. Lewis, D. L., Hagstrom, J. E., Loomis, A. G., Wolff, J. A. & Herweijer, H. Efficient delivery of siRNA for inhibition of gene expression in postnatal mice. Nature Genet. 32, 107–108 (2002). 81. Song, E. et al. RNA interference targeting Fas protects mice from fulminant hepatitis. Nature Med. 9, 347–351 (2003). 82. Hommel, J. D., Sears, R. M., Georgescu, D., Simmons, D. L. & DiLeone, R. J. Local gene knockdown in the brain using viral-mediated RNA interference. Nature Med. 9, 1539–1544 (2003). 83. Kile, B. T. et al. Functional genetic analysis of mouse chromosome 11. Nature 425, 81–86 (2003). 84. Kennerdell, J. R. & Carthew, R. W. Heritable gene silencing in Drosophila using double-stranded RNA. Nature Biotechnol. 18, 896–898 (2000).

NATURE REVIEWS | GENETICS

85. Czauderna, F. et al. Inducible shRNA expression for application in a prostate cancer mouse model. Nucleic Acids Res. 31, e127 (2003). 86. van de Wetering, M. et al. Specific inhibition of gene expression using a stably integrated, inducible smallinterfering-RNA vector. EMBO Rep. 4, 609–615 (2003). 87. Wiznerowicz, M. & Trono, D. Conditional suppression of cellular genes: lentivirus vector-mediated drug-inducible RNA interference. J. Virol. 77, 8957–8961 (2003). Along with references 85, 86, 88, 91 and 92, these authors show that RNAi can be activated in a conditional manner in mammalian systems. 88. Matsukura, S., Jones, P. A. & Takai, D. Establishment of conditional vectors for hairpin siRNA knockdowns. Nucleic Acids Res. 31, e77 (2003). 89. Chen, Y., Stamatoyannopoulos, G. & Song, C. Z. Downregulation of CXCR4 by inducible small interfering RNA inhibits breast cancer cell invasion in vitro. Cancer Res. 63, 4801–4804 (2003). 90. No, D., Yao, T. P. & Evans, R. M. Ecdysone-inducible gene expression in mammalian cells and transgenic mice. Proc. Natl Acad. Sci. USA 93, 3346–3351 (1996). 91. Van Craenenbroeck, K., Vanhoenacker, P., Leysen, J. E. & Haegeman, G. Evaluation of the tetracycline- and ecdysoneinducible systems for expression of neurotransmitter receptors in mammalian cells. Eur. J. Neurosci. 14, 968–976 (2001). 92. Gupta, S., Schoer, R. A., Egan, J. E., Hannon, G. J. & Mittal, V. Inducible, reversible and stable RNA interference in mammalian cells. Proc. Natl Acad. Sci. USA 101, 1927–1932 (2004). 93. Stolarov, J. et al. Design of a retroviral-mediated ecdysoneinducible system and its application to the expression profiling of the PTEN tumor suppressor. Proc. Natl Acad. Sci. USA 98, 13043–13048 (2001). 94. Fraser, A. G. et al. Functional genomic analysis of C. elegans chromosome I by systematic RNA interference. Nature 408, 325–330 (2000). 95. Kamath, R. S. et al. Systematic functional analysis of the Caenorhabditis elegans genome using RNAi. Nature 421, 231–237 (2003). 96. Lee, S. S. et al. A systematic RNAi screen identifies a critical role for mitochondria in C. elegans longevity. Nature Genet. 33, 40–48 (2003). 97. Ashrafi, K. et al. Genome-wide RNAi analysis of Caenorhabditis elegans fat regulatory genes. Nature 421, 268–272 (2003). 98. Pothof, J. et al. Identification of genes that protect the C. elegans genome against mutations by genome-wide RNAi. Genes Dev. 17, 443–448 (2003). 99. Lum, L. et al. Identification of Hedgehog pathway components by RNAi in Drosophila cultured cells. Science 299, 2039–2045 (2003). 100. Kiger, A. et al. A functional genomic analysis of cell morphology using RNA interference. J. Biol. 2, 27 (2003). 101. Boutros, M. et al. Genome-wide RNAi analysis of growth and viability in Drosophila cells. Science 303, 832–835 (2004). 102. Venter, J. C. et al. The sequence of the human genome. Science 291, 1304–1351 (2001). 103. Lander, E. S. et al. Initial sequencing and analysis of the human genome. Nature 409, 860–921 (2001). 104. Waterston, R. H. et al. Initial sequencing and comparative analysis of the mouse genome. Nature 420, 520–562 (2002). 105. Aza-Blanc, P. et al. Identification of modulators of TRAILinduced apoptosis via RNAi-based phenotypic screening. Mol. Cell 12, 627–637 (2003). 106. Zheng, L. et al. An approach to genomewide screens of expressed small interfering RNAs in mammalian cells. Proc. Natl Acad. Sci. USA 101, 135–140 (2004). 107. Shirane, D. et al. Enzymatic production of RNAi libraries from cDNAs. Nature Genet. (2004). 108. Sen, G., Wehrman, T. S., Myers, J. W. & Blau, H. M. Restriction enzyme-generated siRNA (REGS) vectors and libraries. Nature Genet. 36, 183–189 (2004). 109. Berns, K. et al. A large-scale RNAi screen in human cells identifies novel components of the p53 pathway. Nature 428, 431–437 (2004). This work, together with reference 127, describes the generation and applications of array-formatted retrovirus-based barcoded shRNA expression libraries that target human and murine genes. 110. Brummelkamp, T. R., Nijman, S. M., Dirac, A. M. & Bernards, R. Loss of the cylindromatosis tumour suppressor inhibits apoptosis by activating NF-κB. Nature 424, 797–801 (2003). 111. Simmer, F. et al. Genome-wide RNAi of C. elegans using the hypersensitive rrf–3 strain reveals novel gene functions. PLoS Biol. 1, E12 (2003). 112. Ramos-Nino, M. E., Scapoli, L., Martinelli, M., Land, S. & Mossman, B. T. Microarray analysis and RNA silencing link fra-1 to cd44 and c-met expression in mesothelioma. Cancer Res. 63, 3539–3545 (2003).

113. Ruzinova, M. B. et al. Effect of angiogenesis inhibition by Id loss and the contribution of bone-marrow-derived endothelial cells in spontaneous murine tumors. Cancer Cell 4, 277–289 (2003). 114. Harborth, J., Elbashir, S. M., Bechert, K., Tuschl, T. & Weber, K. Identification of essential genes in cultured mammalian cells using small interfering RNAs. J. Cell Sci. 114, 4557–4565 (2001). 115. Elbashir, S. M. et al. Duplexes of 21-nucleotide RNAs mediate RNA interference in cultured mammalian cells. Nature 411, 494–498 (2001). 116. Kapadia, S. B., Brideau-Andersen, A. & Chisari, F. V. Interference of hepatitis C virus RNA replication by short interfering RNAs. Proc. Natl Acad. Sci. USA 100, 2014–2018 (2003). 117. Calegari, F., Haubensak, W., Yang, D., Huttner, W. B. & Buchholz, F. Tissue-specific RNA interference in postimplantation mouse embryos with endoribonucleaseprepared short interfering RNA. Proc. Natl Acad. Sci. USA 99, 14236–14240 (2002). 118. Kawasaki, H., Suyama, E., Iyo, M. & Taira, K. siRNAs generated by recombinant human Dicer induce specific and significant but target site-independent gene silencing in human cells. Nucleic Acids Res. 31, 981–987 (2003). 119. Myers, J. W., Jones, J. T., Meyer, T. & Ferrell, J. E. Jr. Recombinant Dicer efficiently converts large dsRNAs into siRNAs suitable for gene silencing. Nature Biotechnol. 21, 324–328 (2003). 120. Bryom, M., Pallota, V., Brown, D. & Ford, L. Visualizing siRNA in mammalian cells: fluorescence analysis of the RNAi effect. Ambion TechNotes 9, 3–5 (2002). 121. Schwarz, D. S., Hutvagner, G., Haley, B. & Zamore, P. D. Evidence that siRNAs function as guides, not primers, in the Drosophila and human RNAi pathways. Mol. Cell 10, 537–548 (2002). 122. Kirchner, R., Vogtherr, M., Limmer, S. & Sprinzl, M. Secondary structure dimorphism and interconversion between hairpin and duplex form of oligoribonucleotides. Antisense Nucleic Acid Drug Dev. 8, 507–516 (1998). 123. Holen, T., Amarzguioui, M., Babaie, E. & Prydz, H. Similar behaviour of single-strand and double-strand siRNAs suggests they act through a common RNAi pathway. Nucleic Acids Res. 31, 2401–2407 (2003). 124. Shoemaker, D. D., Lashkari, D. A., Morris, D., Mittmann, M. & Davis, R. W. Quantitative phenotypic analysis of yeast deletion mutants using a highly parallel molecular bar-coding strategy. Nature Genet. 14, 450–456 (1996). 125. Ziauddin, J. & Sabatini, D. M. Microarrays of cells expressing defined cDNAs. Nature 411, 107–110 (2001). 126. Kim, D.H. et al. Interferon induction by siRNAs and ssRNAs synthesized by phage polymerase. Nature Biotechnol. 22, 321–325 (2004). 127. Paddison, P.J. et al. A resource for large-scale RNAinterference-based screens in mammals. Nature 428, 427–431 (2004).

Acknowledgements The author thanks the members of his laboratory for critical reading of the manuscript and several other investigators for sharing their unpublished work. We acknowledge that space limitations might have precluded the citation of work of some investigators. The author was supported by grants from Cold Spring Harbor Laboratory and the National Institute of Health.

Competing interests statement The author declares that he has no competing financial interests.

Online links DATABASES The following terms in this article are linked online to: LocusLink: http://www.ncbi.nlm.nih.gov/LocusLink/ AKT1 | CCR5 | CD44 | CXCR4 | cyclophilin B | CYLD | Dicer | DNMT1 | FRA1 | Hh | Id | IGF1R | LMNA | LMO2 | MyoD | PLK1 | RB | SOD1 | TH | TP53 OMIM: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=OMIM X-SCID FURTHER INFORMATION Vivek Mittal’s laboratory: http://www.cshl.org/gradschool/mittal.html Nature Reviews web focus on RNAi: http://www.nature.com/focus/rnai/index.html NCBI Unigene: http://www. ncbi. nlm. nih. gov/BLAST/ RNAi animation: http://www.nature.com/focus/rnai/animations/index.html Access to this links box is available online.

VOLUME 5 | MAY 2004 | 3 6 5