Use of RNA interference libraries to investigate oncogenic signalling in mammalian cells

Article metrics


Over the past decade, ‘RNA interference’ has emerged as a natural mechanism of silencing of gene expression. This ancient cellular antiviral response can be manipulated to provide an effective research tool to knock down the level of expression of selected target genes, providing a very powerful new method for the analysis of cell signalling pathways. Systematic silencing of genes on a genome-wide scale using large rationally designed libraries targeting many thousands of genes provides a novel functional genomics approach to the investigation of many aspects of mammalian cell behaviour, including oncogenic transformation. Here, the different approaches taken to use RNA interference libraries to study the cancer phenotype will be considered, including both selective and high throughput screens and the use of both vector-based and synthetic oligonucleotide-based methods for inducing RNA interference. The advantages and drawbacks of the competing methodologies will be discussed. RNA interference library technology holds great promise for enabling somatic cell genetics in tissue culture systems. Whether it can provide significant new insights into cancer will be its greatest challenge.


RNA interference was first described in plants in 1990 and has subsequently been found as a mechanism of gene silencing in most multicellular eukaryotes. The mechanistic basis for the silencing of gene expression by RNA interference has been described in great detail and has been the subject of numerous recent reviews (Hannon, 2002; Denli and Hannon, 2003; Dykxhoorn et al., 2003; Downward, 2004), so will not be discussed in depth here. In a nutshell, double-stranded RNA molecules in the cell are cleaved by the Dicer enzyme complex to form small interfering (si) double-stranded RNA molecules, which are some 20 base pairs long. These are recognized by another enzyme complex, the RNA-induced silencing complex (RISC), which uses one strand to target complementary mRNA molecules for degradation.

Artificial introduction of siRNA duplexes into cells can thus silence the expression of selected genes. A number of methods have been developed for getting siRNAs of a desired sequence into target mammalian cells. The two most commonly used are to transfect in synthetic short double-stranded RNA molecules of the desired sequence, which then act to engage RISC directly (Elbashir et al., 2001). Alternatively, expression vectors can be introduced into cells, either by transfection or the use of viruses, which direct the production of short hairpin RNA sequences that are then processed into siRNA within the cell by endogenous enzymes (Brummelkamp et al., 2002a, 2002b; Paddison et al., 2002a, 2002b).

Both of these methods of harnessing RNA interference have recently been used on a large scale as a functional genomic tool in mammals, and have been used for longer periods in lower eukaryotes. Systematically silencing genes one by one corresponding to a large fraction of the genome allows the identification of genes that are required for a particular phenotype to occur. In the case of cancer, if genes could be identified in this way, as required for the maintenance of the transformed phenotype, the proteins they encode might make excellent targets for the development of therapeutic drugs. This promise has led to the development of a number of large collections of RNA interference vectors or synthetic oligonucleotides. These are being used in two different types of screens: high throughput and selective. In the former, each assay point corresponds to induction of RNA interference against a single gene, with a phenotypic change being assayed for (Aza-Blanc et al., 2003; Brummelkamp et al., 2003; Paddison et al., 2004). In the latter, many genes are silenced at the same time in a mixed pool of cells, with the screen being designed in such away that only cells acquiring the desired phenotype as a result of knock down of expression of one of these genes can survive: these cells, along with the RNAi sequence they carry are then identified as they emerge at the end of the screen (Berns et al., 2004).

Here the different approaches taken to use RNA interference libraries to study the cancer phenotype will be considered and the advantages and drawbacks of the competing methodologies will be discussed.

RNA interference library design

Oligonucleotide libraries

The principal variable in the design of a large-scale small interfering double-stranded oligoribonucleotide (siRNA) collection is the choice of the sequences used to target each gene. In addition, the level of redundancy with which the library is designed will have major implications, both in terms of cost and the design of experiments that can be performed with it. How many oligos are used to target each gene will depend on the likelihood that any given oligo will successfully knock down its cognate mRNA.

The original criteria for designing siRNA sequences were laid down as a simple set of empirical guidelines by Tuschl in 2001 (Elbashir et al., 2001). Subsequently, a combination of increased understanding of the structure and function of the RISC complex and an empirical analysis of the efficiency of RNA interference induction by very large numbers of siRNA oligos has led to considerable improvements in siRNA design algorithms. Recently, optimized design criteria have been published by a number of groups (Elbashir et al., 2002; Khvorova et al., 2003; Schwarz et al., 2003; Reynolds et al., 2004). Criteria for effective RNA interference include: low G+C content (30–50%), low internal stability at 5′ antisense strand, high internal stability at 5′ sense strand, absence of internal repeats or palindromes, A-form helix between target mRNA and siRNA, presence of A at position 3, U at position 10 and A at position 19 of sense strand, absence of G at position 13 and G or C at position 19 of sense strand, and lack of close homology to other gene sequences.

Publicly available siRNA design algorithms have been created based on these criteria (Cui et al., 2004; Wang and Mu, 2004). A selection of siRNA design algorithms readily available via the internet are: EMBOSS siRNA algorithm (, Promega siRNA Target Designer (, GeneScript siRNA target finder (, Ambion siRNA predictor (, OligoEngine siRNA software (, Whitehead siRNA prediction (, Sfold rational siRNA design (, Hannon Lab RNAi OligoRetriever (, Sonnhammer Bioinformatics Group siSearch siRNA design (

Most publicly accessible siRNA design programs are likely to predict siRNA oligo sequences that will silence gene expression reasonably efficiently (>70% reduction in cognate mRNA) about 50–60% of the time. The use of three such oligos against each gene provides a high probability of success, but does require considerable extra expenditure compared to the use of just one oligo per gene. Recently developed proprietary software from Cenix BioScience has been reported to achieve a >70% gene silencing efficiency with >80% of over 1000 siRNA oligos (, raising the possibility that only one optimally designed oligo per gene could be used to make up an effective genome scale library. As more siRNA sequences are experimentally validated for mRNA knockdown, this approach will become more reliable. In addition to testing siRNA oligos for their effect on expression of endogenous target genes, it is also possible to rapidly screen them against hybrid mRNA fused to a reporter construct (Kumar et al., 2003).

The use of multiple siRNA oligos for each gene raises the issue of whether they should be used singly or as pools with other oligos targeting the same gene. In addition to reducing the number of assays that need to be performed in a screen, an advantage of pooling three or four oligos together is that one maximizes the chance of using an oligo that knocks down expression very effectively. Concerns about the use of pools have focused on the possibility of increasing the likelihood of occurrence of off-target effects. However, as these effects are thought to be dose dependent, if the total oligo concentration applied to the cells remains constant, pools may be less likely to produce significant off-target effects than single oligos. Whether this turns out to be true or not, the importance of verifying effects with a second siRNA sequence against the same gene means that if the library used is only available as pools of oligos, secondary follow-up assays will require the acquisition of individual siRNA oligos, which may incur extra cost and delay.

As RNA interference targets mRNA in the cytoplasm, different siRNAs can be used to target different splice variants transcribed from the same gene. Given the high level of alternative splicing in mammalian genomes, a library targeting all possible splice variants individually would need to be much larger than one targeting common exons.

The high degree of specificity of RNA interference coupled with the considerable differences in sequence between human and rodent genes makes it very unlikely that a library designed against one species would work effectively on the other. Different libraries will need to be designed for each species studied, although it is possible that some degree of crossreaction will be seen with closely related species.

Vector libraries

Similar criteria to the above are used in the design of short hairpin RNA sequences for vector-based RNA interference. However, a critical factor in the success or otherwise of these libraries is the choice of vector backbone. The two large mammalian vector-based RNA interference libraries that have been reported to date use constructs that can be introduced into cells either as retroviral particles or by DNA transfection (Berns et al., 2004; Paddison et al., 2004). Berns et al. used pRETRO SUPER, a retroviral vector in which the sequence encoding the shRNA hairpin is driven by RNA polymerase III acting on the H1 RNA gene promoter (Brummelkamp et al., 2002a). The gene sequence-specific inserts are designed to form hairpins with 19 base pairs of double-stranded RNA, plus a nine nucleotide invariant loop. The vector carries a puromycin-resistance marker. Three sequences were targeted in each of the 7914 human genes.

Paddison et al. used a different retroviral vector, pSHAG MAGIC, in which the sequence encoding the hairpin is driven by RNA polymerase III acting on the U6 RNA gene promoter. In all, 29 base pair hairpins were used, with a four nucleotide loop. A 27-nucleotide U6 RNA leader sequence was also included as this was found to improve knockdown efficiency. The vector also carries puromycin resistance and has a separate unique 60-nucleotide ‘barcode’ built in (see below). To aid in manipulation of the vectors, a recombination system is built into the vector (‘Mating-assisted genetically integrated cloning’) that allows easy exchange of the hairpin sequences between different vector backbones. In all, 9610 human genes and 5563 mouse genes were targeted, mostly with threefold redundancy.

In addition to these two vectors, a wide range of other RNA interference vectors have been designed that could be used for library construction. Incorporation of a fluorescent marker such as GFP could be useful in a library to allow assessment of transfection efficiency: a number of commercial RNAi vectors include this feature, including constructs derived from pRETRO SUPER by Oligoengine, the pSIREN® series from BD Biosciences Clontech, and the GeneSilencer® series from Gene Therapy Systems, Inc. Alternatively, a different viral delivery system could be used, such as adenoviral RNAi vectors (Arts et al., 2003; Zhao et al., 2003), available from BD Biosciences Clontech, Ambion and Invitrogen, or lentiviral vectors (Abbas-Terki et al., 2002; Dirac and Bernards, 2003; Rubinson et al., 2003), available from Invitrogen. Adenoviral vectors are likely to achieve higher levels of expression of the shRNA, while lentiviral vectors will enable expression in nonproliferating cells.

A further future development that could potentially be useful in a vector-based RNA interference library is the incorporation of inducibility. Several tetracycline-inducible RNAi vectors have been reported, both plasmid based (Chen et al., 2003; Czauderna et al., 2003; van de Wetering et al., 2003) and lentivirus based (Wiznerowicz and Trono, 2003). An ecdysone-inducible RNA interference system has also been reported (Gupta et al., 2004). It is not yet clear whether these systems are sufficiently robust to work effectively on a large scale without individual optimization.

RNA interference library use

High throughput screening with oligonucleotide libraries

As mentioned in the introduction, RNA interference library screens can be carried out in a high throughput mode using oligos or vectors, one gene at a time, or in a selective mode using large pools of vectors (Figure 1). High throughput screening has a number of advantages, but also some drawbacks.

Figure 1

Different approaches to large-scale RNA interference screens. See text for details

To screen on a genome-wide scale in high throughput mode can be a daunting prospect, as it requires many thousands of assays to be performed. This is likely to be time consuming and costly. It is essential that screens are designed to minimize these two factors. This requires that assays are highly robust and reproducible, reducing the need for multiple replicates or rescreens. All high throughput RNAi screens performed to date have used some sort of fluorescent readout, which has the advantage of being very easy to measure. It is also important that good positive controls exist to optimize the assays.

Cost considerations also favour assays that can be performed on a very small scale, such as a 96- or 384-well plate format. Aza-Blanc et al. (2003) used a library of 510 siRNA oligos to transfect HeLa cells in 384-well plates and then test them for TRAIL-induced apoptosis, using a cell viability assay with a fluorescent readout. This assay successfully identified several modulators of the apoptotic response, some of which might possibly be misregulated during carcinogenesis. The amount of siRNA oligo transfected was about 10 pmoles/well, meaning that commercially supplied libraries, for example, from Dharmacon, Ambion, or Qiagen, which typically provide 2 nmoles as standard, would in theory be sufficient for about 200 screen runs. While the costs of these commercial libraries may appear prohibitive for academic users, the potential for sharing them between several labs may help alleviate this somewhat. When considering costs, it should also be noted that synthetic oligos and plasmids both require the use of transfection reagents such as lipofection mixes for introduction into cells. Large-scale screening can consume a lot of these expensive consumables and should be factored into any costing of the screen.

Other ways of stretching valuable resources even further can be achieved using reverse transfection in a microarray format. In a series of screens to investigate proteasome function and cytokinesis, both potentially important in carcinogenesis, Silva et al. (2004) spotted siRNA oligos, and also plasmids, in a gelatin solution along with transfection reagent onto glass slides. Cells were overlaid onto the spotted array and took up the siRNAs, showing significant transient knock down of the expression of the target genes and resulting phenotypic changes. In this way, very small amounts of reagents are used and large numbers of assay transfections can be achieved at one time, providing that the end readout can be assessed microscopically. The major limitations of this methodology are that high efficiency of transfection is needed, otherwise spots will need to be very large to find a significant number of transfected cells. This is less of a problem with synthetic siRNA oligos than with plasmid-based shRNA vectors – for which Silva et al. solved the problem by using 293 cells, although these are not ideal for the study of many signalling pathways. The other limitation is that cell motility will restrict how long cells can be followed and how close together elements can be placed.

Once hits have been identified in a high throughput screen, a good secondary screen is needed to check out their significance. As the first assay will ideally give a numerical value, consideration should be given to how big a change from the negative control is needed to be worth pursuing in the secondary screen. This will be influenced by the overall size of the initial screen, the ease with which the secondary screen can be carried out and the level of confidence in the initial results. The secondary screen can be designed to rule out possible off-target effects of the siRNA oligos: for example, using a different sequence against the same gene, or deconvoluting a pool of oligos against the same gene into their individual components.

High throughput screening with vector libraries

Many of the same considerations as those above exist for high throughput screens using vector-based libraries. The major difference is that due to the much lower transfection efficiency of plasmids compared to oligos, it is necessary to mark the transfected cells so that their behaviour can be distinguished from that of the untransfected cells, or to give a measure of the transfection efficiency of the whole cell population in that assay sample for normalization purposes. Paddison et al. (2004) performed a high throughput vector library screen of proteasome function by cotransfecting 293 cells with individual shRNA plasmids, a DsRed fluorescent protein encoding plasmid and an ZsGreen fluorescent protein fused to a PEST sequence containing degron from mouse ornithine decarboxylase. In negative controls, the PEST sequence ensures proteasome-mediated degradation of the ZsGreen fluorescent protein fusion, giving a high ratio of red to green fluorescence. Any shRNA plasmid that leads to proteasome malfunction promotes the levels of ZsGreen and thus reduces the ratio of red to green fluorescence. The screen performed impressively, identifying a high proportion of known proteasomal subunit from a library of 7000 shRNAs tested.

Another screen of this kind performed with relevance to cancer identified the familial cylindromatosis tumour suppressor gene product as a deubiquitinating enzyme important in the regulation of NF-κB (Brummelkamp et al., 2003). This small-scale screen focused on genes with sequence homology to known deubiquitinating enzymes. U2-OS cells were transiently cotransfected with an NF-κB luciferase reporter construct and shRNA vectors individually targeting 50 candidate human deubiquitinating enzymes. The cells were then treated with TNF-α and NF-κB was activity measured by luciferase fluorescence. Targeting the CYLD gene led to upregulation of the NF-κB response, suggesting that the benign tumours commonly seen in familial cylindromatosis patients may be due to increased antiapoptotic signalling by NF-κB.

More complex readouts can be measured in these screens than simply a ratio of fluorescence at two wavelengths. Recently, several companies have developed automated fluorescent microscopy systems, for example, the IN Cell Analyser 1000 and 3000 from Amersham Biosciences and the Discovery-1 from Molecular Devices. These allow, among other things, the automated measurement of the distribution of a fluorescent protein between different cellular compartments, such as the cytosol and the nucleus or the cytosol and the plasma membrane. Such measurements can be performed rapidly in a multiwell format and can add considerable flexibility to the design of high throughput RNAi screens.

A promising approach to screening for potential cancer therapeutic targets using these methods is to look for synthetic lethality between a transforming oncogene and an siRNA. At its simplest, this would involve screening closely related normal and oncogene-transformed cells with an siRNA library looking for cell death, or less optimally arrest, specifically in the untransformed but not the normal cells. Owing to the fact that the microevolutionary process involved in cancer formation tends to result in the cancer cells becoming dependent on oncogenic signalling for their continued survival – a process termed ‘oncogene addiction’ (Weinstein, 2002) – such screens may reveal targets that would allow selective killing of the tumour cells.

Selective screening with vector libraries

The other major assay mode used with RNA interference libraries is the selective screen. A selective pressure is applied that will cause negative control cells to die, growth arrest or in some other way be eliminated from the culture. The only cells that can survive and expand in the culture are those that have received an RNAi sequence that knocks down a gene needed for sensitivity to the selective pressure. The resistant cells can be grown and the sequence of the RNAi sequence determined. The dual constraints of needing relatively long-term selections and having to be able to determine the RNAi sequence present in the resistant cells limits this screening methodology to vector-based RNAi and not synthetic oligos. The final determination of the RNAi sequence in the resistant cells is carried out by PCR amplification and sequencing of the gene-specific insert from invariant primers based on the vector backbone sequence. This can either be carried out on single clones or on populations of resistant cells.

The major example of this type of screen to date was a selection for RNAi sequences that would allow escape of human diploid fibroblasts from p53-mediated senescence (Berns et al., 2004). Human BJ primary fibroblasts expressing telomerase were conditionally immortalized using temperature-sensitive SV40 large T antigen. Upon shift to the nonpermissive temperature, the cells senesce unless they have received an RNAi sequence that blocks this response. Six novel hits were identified that appeared to play a role in p53-mediated senescence.

There are several advantages to the use of selective screens. A major one is the ability to pool large numbers of RNAi vectors together, allowing only a relatively small number of pools to be screened. While this is potentially very useful, these screens have significant limitations. Since the aim of pooling is to achieve a complex mixture of cells each carrying a very small number of RNAi sequences, rather than a homogeneous mixture of cells all carrying a very complex mixture of RNAi sequences, this is only possible with viral-mediated delivery and not with transfection. Viral delivery into human cells normally requires high-level biological containment, but this can be avoided in the case of retroviruses by packaging into an ecotropic delivery system that will only infect mouse cells and using these to screen human cells that have been engineered to express the mouse ecotropic retrovirus receptor (see Berns et al., 2004).

It is hard to assess optimal RNAi pool sizes for these screens. Ideally, a good positive control RNAi vector should be available from knowledge of the system studied. This can then be mixed into increasingly large pool sizes at the same ratio as the other components and tested to see what pool size still allows unequivocal detection of the positive control. A typical pool size used is about a 100 genes; in the case of the NKI library, this was made up of some 300 different retroviruses, allowing for the threefold redundancy (Berns et al., 2004).

Viable pool size will also be strongly dependent on the background level of false positives: high backgrounds will severely limit the possible degree of pooling used and require larger numbers of individual assays. In practice, one of the greatest constraints on this type of screen is provided by the need for low background levels. One approach to easing this is to perform screens iteratively. Ideally this is done by isolating the RNAi sequence inserts from resistant cells, for example, by PCR in the case of the NKI library. This mix of inserts can then be recloned into the RNAi vector and reintroduced into a fresh population of parental cells, which are reselected. This can be repeated a number of times until a limited number of sequences are reproducibly obtained. While this process is time consuming, it can alleviate background problems. There is limited value to reapplying selective pressure to cell populations that have already been selected, as there is no differential selection against the false positives that have already accumulated. Backgrounds can sometimes be reduced by screening a number of single-cell clones of parental cells for low background levels in response to the selective pressure.

As with the high throughput screens, a good secondary screen is very important for verifying hits. Again, this is a good stage at which to check that a different sequence targeted against the same gene can give a positive result, reducing the likelihood of off-target effects. It is also useful to check that in both primary and secondary screens, the induction of a positive phenotype is associated with knockdown of the presumed target mRNA. This also applies to high throughput screens, but is particularly important when using vector-based systems in a stable, rather than transient, manner where knockdown efficiencies tend to be lower. A consequence of this is that the secondary screen here may need to repeat the whole selection protocol to allow cells with a good level of target knockdown to emerge, rather than assuming that all cells receiving the identified RNAi construct will show sufficiently good knockdown to score as phenotypically positive.

A specific problem with selective screens with regard to research into the acquisition and maintenance of the cancer phenotype is that they tend to operate in the reverse direction to that which one would want. By their very nature, cancer cells are able to prosper under circumstances where normal cells do not. For example, they will grow without attachment or without serum. An ideal screen would be to select for loss of the transformed phenotype, but there are very few selections that will efficiently yield up normal cells against a background of transformed cells, whereas many that will do the reverse. One possibility is to screen for genes that when knocked down will promote transformation in the hope that these will be biologically significant tumour suppressor genes. While these are unlikely to provide therapeutic targets directly, they might yield significant prognostic markers or point to the identity of enzymes that reverse their activity, which could be promising cancer drug targets.

Screening using barcodes

A possible way around the difficulty of selecting for a normal cell phenotype from a transformed cell background may be offered by combining selective screens with microarray technology. This strategy, termed ‘barcode’ screening (Brummelkamp and Bernards, 2003; Berns et al., 2004; Paddison et al., 2004), makes use of a gene-specific sequence incorporated into each shRNA vector in the library. In the case of the CSHL library (Paddison et al., 2004), this sequence is separate from the short hairpin sequence, while in the case of the NKI library, it is the short hairpin sequence itself (Berns et al., 2004). Pools of vectors are introduced into cells, which are then selected, for example, by a stress that can result in the loss of cells where a potential therapeutic target gene has been knocked down. In normal screens it would be impossible to determine what sequences were targeted in these cells, as they would be lost to follow-up. However, by amplifying all the barcodes from before and after the selection, labelling them with two different colours fluorescent dyes and hybridizing them to microarrays on which all the barcodes are represented, it is possible to identify genes whose targeting results in loss of cells from the population. This has obvious applications for cancer drug discovery.

In theory, the system could greatly simplify RNAi screening and could be well adapted to screens of relevance to cancer, such as synthetic lethal effects with oncogenes. Berns et al. used it on a small scale to study p53-mediated senescence and were able to identify p53 itself as a positive control. Paddison et al. also tested the bar code system in control experiments, concluding that barcodes that were separate from the hairpin sequences were usable, but that the hairpin itself could not be used effectively. Presumably this is due to the fact that the use of hairpin sequences from the vectors to hybridize against hairpin sequences on the microarray is likely to cause significant problems due to secondary structure, but it is unclear why this worked for Berns et al. but not for Paddison et al. While the barcode strategy clearly holds out significant promise for the future, it is not yet a fully validated technology and no novel findings have yet been reported using it.

Problems with RNA interference library methodology

A number of problems that arise in RNA interference library screens have been raised above. Particularly, critical problem areas or issues not yet addressed are:

Redundancy of target genes

As a genetic approach, RNAi library screening may be limited by redundancy of the target genes. This is a much greater problem in mammalian systems than in organisms such as worms and flies, where a great deal of RNAi screening has been carried out already. In the case of high throughput screens, judicial pooling of RNAi oligos or constructs against two or three very closely related proteins might alleviate this. However, in selective screens using large viral pools, this is unlikely to be workable as each cell takes up only a very small number of viruses from the complex pool and the chances of getting the desired combination would be very low.

Low efficiency of knock down

As the algorithms for selecting optimal sequences for RNAi improve, this problem may recede. This is a particular concern with vector-based RNAi, which tends to be less efficient than synthetic oligonucleotide-driven silencing. Stable expression of RNAi vectors tends to be still less efficient at silencing than transient, making this problem the greatest for selective screens. In selective screens using complex viral mixtures, it is highly likely that cells will only carry one integrated copy of any given construct, so this must be capable of providing an effective knock down if the screen is to succeed. This is the most demanding situation imaginable for achieving gene silencing. However, if the background in a selective screen is sufficiently low, it may be possible to identify rare cells in which the vector is integrated in a particularly favourable location, resulting in a much better knockdown than average for that construct.

Problems handling large collections of constructs

There is a particular difficulty dealing with large numbers of plasmids that is not an issue with oligonucleotide collections. Preparing transfectable DNA from large numbers (thousands) of plasmids for high throughput vector library screens is time consuming and costly. It also subjects the bacteria carrying the plasmids to strong selective pressure that can result in recombination events. Replicating multiple copies of a gridded library can run into similar problems. Both the NKI and CSHL RNAi libraries have suffered from problems with recombination resulting in loss of hairpin sequences. There is also a problem when using pooled vectors that the complexity of the pool is reduced when the bacteria are grown. It is likely that even if the originally produced libraries are exactly as intended, by the time they have been distributed and manipulated by multiple users, they will be significantly less than optimal.

Off-target effects

There are conflicting views as to how large a problem this is. Two forms of off-target effect exist: sequence specific and sequence independent. The former is caused by siRNAs targeting expression of genes with related sequence to the desired target. Systematic studies using microarrays have concluded either that this is a significant problem, with similarities as low as 11 contiguous bases causing offtarget effects (Jackson et al., 2003), or that it is not a significant problem (Chi et al., 2003). While the former study seems likely to be an overly pessimistic view, it does underscore the importance of making sure that hits from the screens are not due to crossreaction, most readily by using different sequences targeting the same gene in a secondary screen and in the subsequent follow-up. All hits from these screens require extensive follow-up, so in practice this is not much of an added burden. While the studies cited above studied mRNA degradation, it is far less clear how much off-target activity might result from inhibition of mRNA translation by imperfectly matched sequences, so-called micro RNA effects (He and Hannon, 2004).

The major concern with sequence-independent effects is the induction of interferon responses (Bridge et al., 2003). It is still unclear how significant this problem is, but it is relatively simple to incorporate an assay for interferon induction as part of follow-up screens, for example quantitative PCR for expression of an interferon-induced gene such as OAS1. All off-target effects will be concentration dependent, so it is important to use no more siRNA agent than necessary, especially in transient expression screens.

Conclusions and future prospects

The use of RNA interference in mammalian cells is only 3 years old, but it has already been scaled up such that the function of very significant fractions of the genome can be interrogated using RNA interference libraries. Several groups have been quick to see the potential of this approach to the discovery of novel components of signalling pathways important in cancer, some of which may prove to be useful new drug targets. While some important findings have already been reported, it is clear that the methodology used so far is in its infancy, although it is developing very fast. The two large mammalian vector-based RNA interference libraries reported to date have significant limitations and probably should be viewed only as prototypes, but will no doubt lead to the development of new generation libraries with much improved efficiency. Rapid improvements in our understanding of what sequences are likely to induce effective knockdown of gene expression will greatly improve the effectiveness of both vector- and oligonucleotide-based libraries. In addition, improved assay technology, in particular, involving developments in fluorescence microscopy, is also greatly strengthening the power and versatility of these screens. Of the two assay strategies, high throughput and selective, high throughput screening is clearly the most versatile, and can be adapted to investigate a very wide range of biological phenomena. In particular, it can readily be used to study several of the major characteristics associated with cancer. By contrast, the potential use of selective screens is much more limited due to the requirement for a stringent elimination of all cells apart from those with the desired phenotype during the course of the assay. This can be very challenging to achieve and would rule out straightforward assays for reversion for transformation. However, when successfully designed, they can deliver novel components of pathways that may be important in cancer, especially in restraint of transformation. It is clear that the near future will see a great increase in the use of the RNA interference library approach to study cancer and may well start to uncover large numbers of potential new therapeutic targets that would have taken a very long time to find by previous methodology.


  1. Abbas-Terki T, Blanco-Bose W, Deglon N, Pralong W and Aebischer P . (2002). Hum. Gene. Ther., 13, 2197–2201.

  2. Arts GJ, Langemeijer E, Tissingh R, Ma L, Pavliska H, Dokic K, Dooijes R, Mesic E, Clasen R, Michiels F, van der Schueren J, Lambrecht M, Herman S, Brys R, Thys K, Hoffmann M, Tomme P and van Es H . (2003). Genome. Res., 13, 2325–2332.

  3. Aza-Blanc P, Cooper CL, Wagner K, Batalov S, Deveraux QL and Cooke MP . (2003). Mol. Cell, 12, 627–637.

  4. Berns K, Hijmans EM, Mullenders J, Brummelkamp TR, Velds A, Heinerikx M, Kerkhoven RM, Madiredjo M, Nijkamp W, Weigelt B, Agami R, Ge W, Cavet G, Linsley PS, Beijersbergen RL and Bernards R . (2004). Nature, 428, 431–437.

  5. Bridge AJ, Pebernard S, Ducraux A, Nicoulaz AL and Iggo R . (2003). Nat. Genet., 34, 263–264.

  6. Brummelkamp TR and Bernards R . (2003). Nat. Rev. Cancer, 3, 781–789.

  7. Brummelkamp TR, Bernards R and Agami R . (2002a). Cancer Cell, 2, 243–247.

  8. Brummelkamp TR, Bernards R and Agami R . (2002b). Science, 296, 550–553.

  9. Brummelkamp TR, Nijman SM, Dirac AM and Bernards R . (2003). Nature, 424, 797–801.

  10. Chen Y, Stamatoyannopoulos G and Song CZ . (2003). Cancer Res., 63, 4801–4804.

  11. Chi JT, Chang HY, Wang NN, Chang DS, Dunphy N and Brown PO . (2003). Proc. Natl. Acad. Sci. USA, 100, 6343–6346.

  12. Cui W, Ning J, Naik UP and Duncan MK . (2004). Comput. Methods Prog. Biomed., 75, 67–73.

  13. Czauderna F, Santel A, Hinz M, Fechtner M, Durieux B, Fisch G, Leenders F, Arnold W, Giese K, Klippel A and Kaufmann J . (2003). Nucleic Acids Res., 31, e127.

  14. Denli AM and Hannon GJ . (2003). Trends Biochem. Sci., 28, 196–201.

  15. Dirac AM and Bernards R . (2003). J. Biol. Chem., 278, 11731–11734.

  16. Downward J . (2004). BMJ, 328, 1245–1248.

  17. Dykxhoorn DM, Novina CD and Sharp PA . (2003). Nat. Rev. Mol. Cell. Biol., 4, 457–467.

  18. Elbashir SM, Harborth J, Lendeckel W, Yalcin A, Weber K and Tuschl T . (2001). Nature, 411, 494–498.

  19. Elbashir SM, Harborth J, Weber K and Tuschl T . (2002). Methods, 26, 199–213.

  20. Gupta S, Schoer RA, Egan JE, Hannon GJ and Mittal V . (2004). Proc. Natl. Acad. Sci. USA, 101, 1927–1932.

  21. Hannon GJ . (2002). Nature, 418, 244–251.

  22. He L and Hannon GJ . (2004). Nat. Rev. Genet., 5, 522–531.

  23. Jackson AL, Bartz SR, Schelter J, Kobayashi SV, Burchard J, Mao M, Li B, Cavet G and Linsley PS . (2003). Nat. Biotechnol., 21, 635–637.

  24. Khvorova A, Reynolds A and Jayasena SD . (2003). Cell, 115, 209–216.

  25. Kumar R, Conklin DS and Mittal V . (2003). Genome Res., 13, 2333–2340.

  26. Paddison PJ, Caudy AA, Bernstein E, Hannon GJ and Conklin DS . (2002a). Genes Dev., 16, 948–958.

  27. Paddison PJ, Caudy AA and Hannon GJ . (2002b). Proc. Natl. Acad. Sci. USA, 99, 1443–1448.

  28. Paddison PJ, Silva JM, Conklin DS, Schlabach M, Li M, Aruleba S, Balija V, O’Shaughnessy A, Gnoj L, Scobie K, Chang K, Westbrook T, Cleary M, Sachidanandam R, McCombie WR, Elledge SJ and Hannon GJ . (2004). Nature, 428, 427–431.

  29. Reynolds A, Leake D, Boese Q, Scaringe S, Marshall WS and Khvorova A . (2004). Nat. Biotechnol., 22, 326–330.

  30. Rubinson DA, Dillon CP, Kwiatkowski AV, Sievers C, Yang L, Kopinja J, Rooney DL, Ihrig MM, McManus MT, Gertler FB, Scott ML and Van Parijs L . (2003). Nat. Genet., 33, 401–406.

  31. Schwarz DS, Hutvagner G, Du T, Xu Z, Aronin N and Zamore PD . (2003). Cell, 115, 199–208.

  32. Silva JM, Mizuno H, Brady A, Lucito R and Hannon GJ . (2004). Proc. Natl. Acad. Sci. USA, 101, 6548–6552.

  33. van de Wetering M, Oving I, Muncan V, Pon Fong MT, Brantjes H, van Leenen D, Holstege FC, Brummelkamp TR, Agami R and Clevers H . (2003). EMBO Rep., 4, 609–615.

  34. Wang L and Mu FY . (2004). Bioinformatics, 20, 1818–1820.

  35. Weinstein IB . (2002). Science, 297, 63–64.

  36. Wiznerowicz M and Trono D . (2003). J. Virol., 77, 8957–8961.

  37. Zhao LJ, Jian H and Zhu H . (2003). Gene., 316, 137–141.

Download references

Author information

Correspondence to Julian Downward.

Rights and permissions

Reprints and Permissions

About this article


  • RNA interference
  • screening
  • transformation
  • library
  • oncogenesis

Further reading