The evolution of RNAi as a defence against viruses and transposable elements

1Broadly defined, RNA interference (RNAi) constitutes a class of processes that use short RNAs (approx. 20–30 nucleotides) to recognize and manipulate complementary nucleic acids. RNAi-related pathways have roles in the control of gene expression, epigenetic modification and regulation of heterochromatin, and in the host–parasite interactions (see §2 for references). RNAi-related phenomena are so pervasive that it now seems surprising that the molecular basis of RNAi was first reported little more than 15 years ago (more that 97% of papers using the terms ‘RNAi’ and ‘RNA silencing’ have been published since 2001; for a historical perspective see ). From a host–parasite and host–pathogen viewpoint, RNAi-related pathways provide a particularly interesting arena for coevolutionary interaction. For the host, RNAi is a defence against both viruses and ‘genomic parasites’ such as transposable elements (TEs), and can provide fine control over other components of innate immunity by modulating gene expression. For the parasite, host RNAi can be exploited to subvert host cell function by manipulating host gene expression and blocking host resistance mechanisms. Strikingly, components of RNAi are widely distributed across the eukaryotic phylogeny, implying that the common ancestor of all eukaryotes had a functional RNAi pathway, over a billion years or more ago.

2Immune system genes and the parasite molecules they interact with often evolve rapidly. It is thought that this results from a coevolutionary arms race, in which there is a reciprocal process of adaptation and counteradaptation between parasites and hosts. As RNAi is a primary defence of many organisms against viruses and TEs, it is likely to be a key battlefield on which these arms races are played out. This has the potential to drive the rapid evolution of the proteins involved, and ultimately shape the RNAi pathways themselves.

3The major RNAi mechanisms have been recently reviewed (see §2) and it is not our purpose to duplicate that effort here. Instead, following a brief outline of the pathways most relevant to host–parasite biology, we review RNAi-mediated host–parasite interactions from an evolutionary perspective. We also present new analyses of the distribution of key RNAi-pathway genes across the animal phylogeny, and the rates of evolution in both RNAi genes and viral genes that suppress RNAi.

4The RNAi-related pathways are unified by their dependence on sequence-specific binding between short (approx. 20–30 nt) RNAs and target sequences, and their common use of Argonaute family proteins. However, this belies an enormous underlying diversity between pathways and phylogenetic lineages. For example, the short RNA molecules can be derived from exogenous double-stranded RNA (dsRNA) (including viral genomes and replicative intermediates), fold-back RNAs expressed by the host or virus, or overlapping pairs of sense and antisense transcripts. Different pathways also use different complements of Dicer, Argonaute and accessory proteins, and the outcome of RNAi can be the cleavage and degradation of target RNA, the recruitment of additional factors to modify gene expression, or even long-term expression changes through epigenetic modification and heterochromatin formation. (For recent reviews see;;;;;;; ). For a glossary of relevant terms see table 1.

5The best-studied role of RNAi in the immune system is as a direct defence against viruses (figure 1a; reviewed by ). dsRNA from viruses is recognized by Dicer and cut (‘diced’) into short 21–24 nucleotide fragments called short interfering RNAs (siRNAs, also known as viRNAs when they are derived from viruses; ). These are then loaded into an Argonaute-containing effector complex (RISC, the RNA-induced silencing complex; figure 1), and one strand of the viRNA is cleaved and degraded (reviewed by ). The active Argonaute complex then cleaves (‘slices’) viral RNA with the complementary sequence to the viRNA (figure 1; ).

6In addition to this core siRNA pathway (figure 1a), many eukaryotes are able to amplify the viRNA pool using a host-encoded RNA-dependent RNA polymerase (RdRp), and propagate the antiviral defence signal to other cells. In Caenorhabditis elegans, a host RdRp binds the siRNA–target duplex and directly synthesizes secondary siRNAs, reportedly independently of additional Argonaute or Dicer activity. In plants, a host RdRp is also recruited by the siRNA–Argonaute complex, but in this case it synthesizes long dsRNAs which are then processed into siRNAs by a Dicer family protein. This is thought to amplify viral-derived sequences in both plants and nematodes, and in plants RdRp knockout mutants are more susceptible to viral infection.

7In plants and nematodes, RNAi initiated in one cell can be propagated to neighbouring cells, and even to more distant tissues (‘non-cell-autonomous RNAi’; ). In C. elegans, the movement of siRNAs between cells depends on SID-1 (systemic RNAi defective 1), which is thought to be a passive channel for cell-to-cell RNA movement. In Arabidopsis, 21 nt siRNAs are transported from cell to cell for short distances (10–15 cells), and an unknown signal (probably dsRNA) is loaded into the phloem to provide long-distance transport of the silencing signal. It has been argued that the short RNAs are likely to need to be amplified by an RdRp for the RNAi signal to be effectively propagated between cells, and in Arabidopsis movement beyond 15 cells is indeed dependent on RdRp function. However, many animals encode SID-1 homologues but not RdRps (see §6), and there is evidence of non-cell-autonomous RNAi in several taxa that lack their own RdRp.

8The microRNA (miRNA; figure 1b) pathway is a vital component of post-transcriptional control of gene expression in plants and animals. MicroRNAs are encoded by the host genome and transcribed by RNA polymerase II as part of larger fold-back transcripts (primary miRNAs), which are processed in the nucleus by Drosha family members to form short stem-loops (pre-miRNAs), and then exported to the cytoplasm for processing by a Dicer family member to form the mature miRNA (figure 1b, reviewed by;; ). As with viRNAs, miRNAs are loaded into an Argonaute-containing effector complex (RISC), where the complementary ‘passenger’ strand is lost. This miRNA effector complex can then interact with target messenger RNAs that contain complementary sequence to the miRNA. In plants, miRNAs are usually an exact match to their mRNA targets and cause the cleavage and degradation of the target transcript. In animals, miRNAs usually pair imperfectly with one or more targets in the 3′-UTR and do not induce mRNA endonucleolytic cleavage, but inhibit translation by a variety of mechanisms. Although historically considered as a mechanism for the host to control its own gene expression, the miRNA pathway has more recently been shown to be important in interactions with viruses, with both hosts and viruses producing miRNAs that mediate host–virus interactions (see §4).

9A third group of RNAi-related pathways target transposable elements (TEs: transposons, retrotransposons and endogenous retroviruses;;;;;;;;; ). It was observed in as early as 1999 that several RNAi mutants of C. elegans had elevated rates of transposition, and it is now apparent that RNAi also plays an important role in determining the activity of TEs in many other eukaryotes. TE transcripts are processed by Argonaute and/or Dicer family members, both reducing transcript numbers and leading to epigenetic modification of genomic copies of the transposons, heterochromatin formation and reduced expression of TEs.

10In Arabidopsis, the pathway that targets transposons involves Dicer-like 3 and Argonaute 4, and a class of 24–26 nucleotide siRNAs derived from the TE transcripts. In Drosophila and vertebrate germ lines, TE silencing relies on Argonaute proteins in the Piwi family, and a class of short RNAs known as Piwi-interacting short RNAs (piRNAs;;; ). These are longer (24–30 nt) than viRNAs and miRNAs, and originate from a single-stranded RNA (ssRNA) precursor without the involvement of a Dicer protein. In Drosophila, the majority of piRNAs target repeat sequences such as TEs (also known as repeat-associated siRNAs or rasiRNAs), while in vertebrates most target other repetitive sequences and only a minority are complementary to TEs. A model has been proposed in which the suppression of TEs is maintained by a cyclic feedback process that alternately cleaves sense and antisense TE transcripts (‘ping-pong’: figure 1c; see for a review).

11Very recently, it has also been shown that Drosophila and mice have an additional RNAi pathway that targets endogenous dsRNA such as that produced by overlapping 3′-UTRs from genes on opposite strands, long hair-pin fold-back sequences and TEs. In Drosophila, this pathway is similar to the antiviral pathway (figure 1a)—requiring both Ago-2 and Dcr-2—but it differs in that Loqs (from the miRNA pathway; figure 1b) takes the role of R2D2, at least for some transcripts. The resulting ‘endogenous siRNAs’ have similar size and properties to viral-derived siRNAs (viRNAs), and have a role in regulating the expression of several host genes in addition to affecting TE transcript levels.

12Most organisms lack the adaptive immune system of vertebrates, and must rely on innate immunity to protect themselves from infection. The innate immune system often detects and destroys pathogens by recognizing families of molecules that are conserved across broad classes of pathogens. In the case of bacteria and fungi, these molecules are primarily cell-surface polysaccharides, while viruses can be recognized if they produce dsRNA. dsRNA can be used to distinguish self from non-self because eukaryotes typically do not produce long stretches of dsRNA (other than that destined for processing by RNAi). However, many RNA viruses have double-stranded genomes, and even viruses with single-stranded genomes can produce dsRNA during replication, when viral RNA forms secondary structures, or when there are complementary sense and antisense replicative intermediates. DNA viruses can potentially also produce the dsRNA necessary for RNAi through secondary structure in their transcripts or sense–antisense transcript pairs. It is now clear that the viRNA pathway (figure 1a) is an important component of innate antiviral immunity in plants, fungi, arthropods, nematodes and many animals.

13The possibility that RNAi might have an antiviral function was first raised in plants when experimentally induced ‘gene silencing’ was found to provide resistance to viruses carrying an identical sequence. It was then identified as a natural component of innate antiviral immunity when viruses were found to naturally induce a similar response. For animals, antiviral RNAi was first identified in Drosophila cell culture, where the beetle Nodamura virus FHV (flock house virus) acts as both an initiator and a target of the viRNA pathway. It was later confirmed as biologically relevant defence outside of cell culture using the natural host–virus combination of O'nyong-nyong virus (alphavirus; Togaviridae) and Anopheles gambiae mosquitoes. RNAi has subsequently been identified as a component of antiviral immunity in adult Drosophila and in nematode worms. Very recently, RNAi has also been identified as an important antiviral defence in fungi. Therefore, although several isolated lineages lack components of RNAi (e.g. Saccharomyces), it is likely that the viRNA antiviral pathway is a shared character of most eukaryotes.

14It remains an open question as to whether this pathway might be a component of antiviral immunity in vertebrates. Although vertebrates encode relevant genes (including Argonaute, Dicer and SID family members; figure 5), there is no strong evidence of a viRNA antiviral pathway, and these RNAi genes are used in other RNAi pathways (e.g. miRNA function and anti-TE function) that mediate host cellular processes.

15Viral suppression of the immune system is a widespread phenomenon (e.g.; ). It is therefore not surprising that many viruses inhibit the viRNA pathway by expressing viral suppressors of RNAi (VSRs;; ). Very soon after the natural antiviral role of RNAi was revealed, the Hc-Pro protein of plant Potyviruses and the 2b protein of plant cucumoviruses were identified as VSRs. The first VSR identified in an animal virus was the B2 protein of the beetle virus FHV, which was initially found to suppress RNAi in plants, before being shown to play the same role in insects. As of 2006, Li and Ding were able to review the function of more than 50 VSRs from over 30 viral genera. This list includes all categories of RNA viruses (negative sense, positive sense and double stranded) and some DNA viruses, in both plants and animals. Although the VSR function of some of these genes has been questioned (especially those of negative-sense ssRNA animal viruses, e.g. ), it now appears that active suppression of the host viRNA pathways is often an important component of infection, and nearly universal in plant–RNA virus interactions. VSRs appear so important for successful viral infection that several viruses encode multiple VSRs, e.g. Citrus Tristeza Virus (a Closterovirus) encodes three and several potyviruses encode two.

16Many VSRs are dsRNA-binding proteins. However, while sequestering viRNA molecules away from the RNAi pathway is apparently the primary mechanism for some—such as P19 of tombusviruses, which shows a strong specificity for 21 nt dsRNA duplexes—it is now clear that there are a diverse range of other suppression mechanisms. Indeed, not all VSRs are even proteins—at least two are thought to be RNAs, which may function by binding RNAi components in place of viRNAs. Different VSRs also target different components of RNAi. For example, both Hc-Pro and B2 appear to inhibit viRNA processing from long dsRNA precursors by Dicer. Conversely, the 2b protein of cucumoviruses and the P0 protein of poleroviruses both target Argonaute family members; 2b by binding Arabidopsis Ago-1 directly to prevent the RISC complex from cleaving its target RNA, and P0 by targeting Argonaute family proteins for degradation.

17In plants, where cell-to-cell propagation of RNAi is a key feature of the viRNA pathway (and where assays for non-cell-autonomous silencing are best developed), VSRs differentially affect cell-autonomous RNAi, non-cell-autonomous RNAi and the vascular transport of RNAi. For example, if the P38 VSR is deleted from Turnip Crinkle Virus (Tombusviridae), the virus is still able to move locally from cell to cell, but cannot spread through the entire plant. Similarly, the Potexvirus VSR P25 does not effectively inhibit cell-autonomous RNAi, but does block long distance movement. By contrast, Hc-Pro from potyviruses and AC2 from African cassava mosaic virus (a DNA geminivirus) are VSRs that suppress cell-autonomous and short-distance RNAi, respectively, but not long-distance export. A second strategy may be for a virus to manipulate the host's own control of the pathway rather than suppressing RNAi directly. For example, the Begomovirus VSR (AC2) appears to act as a transcriptional regulator that alters the expression of approximately 50 host genes, including some that might modify RNAi activity.

18Viruses may also reduce the effects of RNAi by preventing dsRNA from coming into contact with the RNAi pathway, or avoiding degradation once viRNAs have been produced. Viruses have several adaptations to reduce the production of dsRNA and shield the viral genome from degradation. For example, when single-stranded viruses replicate they must produce the complementary replication intermediate, and the nucleocapsid protein of negatively stranded viruses prevents the two strands from annealing and producing dsRNA. Because viRNAs are rarely derived uniformly from across the viral genome (e.g.; ), it seems likely that certain regions of the genome are more vulnerable than others. The secondary structure of RNA can determine which fragments are processed to form viRNAs, and whether the active silencing complex has access to degrade the target RNA. For example, although viroids (subviral RNA pathogens that do not encode proteins) are a source of functional viRNAs, they are not cleaved by the silencing complex owing to their secondary structure. When HIV is cultured with an artificial RNAi construct, its secondary structure evolves rapidly to block access by RNAi components, supporting the idea that this could be a viable mechanism of viral evasion, at least in principle.

19There is clearly an evolutionary conflict between the viRNA pathway and its suppression by viruses. This has the potential to result in a never-ending arms race, where the RNAi pathway continually evolves new ways to escape suppression by VSRs, which leads to counteradaptations by the virus that restore suppression. If this is the case, then VSRs might be expected to evolve much faster than other viral genes. This is certainly true when comparing VSRs across different viral families. observed that VSRs from different viral families have essentially no structural similarity, even where they have similar functions. This is particularly striking when the dsRNA-binding motifs are compared across different VSRs: the dsRNA-binding domains from P19 (tombusviruses), P21 (Beet Yellows Virus) and B2 (FHV) each appear to have evolved different and entirely novel dsRNA-binding motifs (reviewed by ). Many VSRs also appear to have arisen relatively recently, and it has been argued that this is an indication of frequent viral adaptation to host RNAi.

20If VSRs are engaged in an evolutionary arms race, then we might also expect to see the rapid sequence evolution of homologous VSRs within the same virus family. There is circumstantial evidence that this is the case. For example, the P1 VSR of potyviruses (Valli et al., ) is highly variable both in length and sequence, and shows evidence of recombination between potyviral lineages. Potyviruses typically have a second VSR called Hc-Pro, but in cucumber vein yellowing virus from the related genus Ipomovirus, Hc-Pro has been lost and its function appears to have been taken on by a tandemly duplicated copy of P1. Similarly, in the Closteroviridae there is a considerable sequence diversity in the VSR P22, and in one species (Sweet Potato Chlorotic Stunt Virus) some viral isolates have lost this suppressor altogether but still maintain VSR activity, presumably from a second locus.

21To test whether VSRs evolve faster than other viral genes, we analysed the rate of protein divergence for all genes in pairs of isolates from 17 ssRNA plant viruses with experimentally verified VSRs (figure 2). We found that protein divergence (relative to silent site divergence) between pairs of isolates was significantly higher for VSRs than for other genes, which is consistent with strong selection acting on VSRs. However, we would caution that the rapid evolution of VSRs could also be explained by low selective constraints (if most of the amino acid changing mutations have little effect on the function of the protein, they may be fixed at a high rate by random genetic drift), and more sophisticated analyses are needed to reject this possibility.

22There is evidence that some viruses might have evolved to subvert the viRNA pathway for their own benefit (reviewed by ). Several short RNAs derived from Cauliflower Mosaic Virus are complementary to Arabidopsis messenger RNAs, and downregulate those genes. This may be coincidental rather than adaptive—since Arabidopsis is unlikely to be a natural host for cauliflower mosaic virus, this would require the short RNA targets to be very highly conserved between members of the Brassicaceae—but again illustrates the potential for viral evolution in response to host RNAi. It opens up interesting opportunities for the virus, as these host-manipulating viRNAs could in principle be exported from the site of infection and ‘prime’ the rest of the plant for viral invasion.

23Because many viruses rely on suppressing the viRNAi pathway to infect their hosts, there will be a considerable advantage to mutations in viRNAi genes that escape this suppression. Furthermore, as the VSRs themselves are evolving quickly, the viRNAi pathway faces a continually changing array of suppressors to adapt to, and this could result in an evolutionary arms race driving the rapid evolution of both host and viral proteins. As predicted by this hypothesis, three key proteins in the viRNAi pathway of Drosophila (Dcr-2, Ago-2 and R2D2) are among the top 3 per cent of the most rapidly evolving in the entire genome (; figure 3). Proteins involved in the immune system often have a higher rate of evolution than the genome average, but even compared with other immunity genes the viRNAi genes evolve exceptionally fast. Indeed, no other functional class of immunity genes shows such consistently rapid evolution. These three viRNA genes have paralogs in the miRNA pathway (Dcr-1, Ago-1 and R3D1/Loqs; figure 1), which have similar molecular functions but no direct antiviral function. The miRNA-pathway genes evolve slowly, suggesting that it is the antiviral role of viRNAi genes that is causing the high rate of evolution.

24However, rapid protein evolution need not be caused by positive natural selection, but can also result from low selective constraints. If a gene is evolving neutrally, the ratio of non-synonymous (amino acid changing) to synonymous differences between species will be the same as the ratio for polymorphism within species. In the case of the viRNAi genes, the high amino acid divergence between species is not matched by the high levels of amino acid polymorphism, providing compelling evidence that selection has driven the rapid evolution of these genes. Using this excess of amino acid divergence relative to polymorphism, we have estimated the number of amino acid differences between Drosophila melanogaster and its close relative Drosophila simulans that were fixed because they had a selective advantage (figure 4). The rate at which natural selection has fixed selectively advantageous amino acid substitutions in the viRNAi genes is 24-fold greater than their miRNA paralogs. For R2D2 and Dicer-2 these results have also been confirmed in a second study, using a different approach that compared the genome sequences of 12 species of Drosophila.

25It has recently become clear that the miRNA pathway can also play a role in host–virus interactions. Unlike the viRNA pathway, which primarily uses short RNAs derived from viruses as an immune defence, both host and virus sequences can be processed into miRNAs, and the miRNA pathway is not only antiviral but can also be exploited by viruses for their own benefit.

26Virus-encoded miRNAs were first identified in the Epstein–Barr virus and have now been found in at least 12 mammalian viruses, spanning the three herpesvirus subfamilies, polyomaviruses and retroviruses (for a current listing see the miRBase Sequence Database; ) (for a review see ). The miRNA pathway (figure 1b) provides a powerful mechanism for fine tuning gene expression in the host, and viruses make use of this pathway to modulate their own gene expression. For example, in Simian Virus 40 (SV40), the viral miRNA miR-S1 is expressed late in infection and downregulates an early viral gene that is encoded on the opposite strand of the miRNA. This autoregulation of the viral gene (the viral T-antigen) by an miRNA is proposed to be important for enabling the virus to evade the host immune response.

27It has been estimated that more than 50 per cent of the genes in herpesviruses and poxviruses could be devoted to manipulating the host immune system and viral miRNAs are an efficient way to do this as they only require a small amount of coding sequence in the genome, yet can have a large capacity for gene regulation as they can have multiple targets. This field is in its infancy and the actual number of targets and in vivo functions of virus-encoded miRNAs remains to be elucidated. However, a recent report demonstrates that a miRNA encoded by Human Cytomegalovirus (HCMV) can target the MHC class I-related chain B gene and thereby protect HCMV-infected cells from lysis by natural killer cells. Kaposi's sarcoma-associated herpesvirus appears to use a different strategy to manipulate host function: it encodes a miRNA (termed mir-k12-11) that mimics the host-encoded miRNA mir-155, both having the same sequence in the seed region and sharing several targets. Some of these genes regulate cell growth and B-cell transformation, so the miRNA may be partly responsible for B-cell tumours in infected patients.

28Given the apparent benefits to the virus, it might be expected that many viral lineages will be found to encode miRNAs. Of the 12 viruses reported to encode miRNAs, 10 belong to the herpesvirus family, one belongs to the polyomavirus family and one to the retrovirus family. This bias is at least partly attributable to the large volume of research aimed at herpesviral miRNAs and additional sequencing efforts will be required to understand the generality of viral-miRNA existence. However, computational analyses reveal a low probability that pre-miRNA structures occur in human and mouse ssDNA and ssRNA viruses, and cloning efforts failed to identify miRNAs in RNA viruses including Yellow Fever Virus and Hepatitis C Virus. Indeed, with the exception of miRNAs recently reported from HIV, viral miRNAs have only been found in dsDNA viruses. Why might this be? First, the primary miRNA-processing machinery (e.g. Drosha) is found in the nucleus, and thus is unlikely to be available to cytoplasmic viruses. Second, host-targeted viral miRNAs must match the host sequences, and may only have time to evolve where there is a long-term host–parasite association. Third, some viruses produce VSRs that can inhibit the miRNA pathway. These factors may explain the apparent bias for miRNA representation in herpesviruses, which replicate in the nucleus, tend to remain associated with the same host lineage for millions of years, and are not known to inhibit the miRNA pathway.

29Hosts can also use the miRNAs pathway as a defence against viruses by expressing miRNAs that target viral RNA. For example, Primate Foamy Virus 1 (PFV-1) mRNA is targeted by the human miRNA mir-32, which results in decreased PFV-1 replication. However, the physiological relevance of this might be questioned, since mir-32 may not be expressed in cell types relevant to PFV-1 replication, foamy viruses currently only occur as rare zoonoses in humans and it is not known whether the target sequence is conserved in related viruses that regularly infect humans. In another recent report, mice deficient in Dicer-1 were shown to be more susceptible to infection by Vesicular Stomatitis Virus (VSV): two cellular miRNAs, mir-24 and mir-93, are predicted to target VSV genes encoding the viral RdRp and a polymerase cofactor, and mutation of the miRNA-binding sites in these genes rescues the growth defect in Dicer-1 deficient cells. It has also been suggested that miRNAs may be an effector in the classical vertebrate innate immune response. Activation of the interferon beta pathway in human cells infected by Hepatitis C Virus induces the expression of several host-encoded miRNAs, some of which have predicted targets in the virus. In support of their antiviral role, when these miRNAs are experimentally introduced they reproduce the antiviral effects of interferon beta on HCV, and the interferon defence is lost when they are experimentally removed. It appears that host miRNAs can also modulate cellular genes involved in the interferon response, as demonstrated for mir-146, and the expression of mir-146 is induced by the EBV-encoded latent membrane protein (LMP1), which suggests an intricate role of miRNAs in viral–host interactions.

30Viruses may also exploit host-encoded miRNAs for their own advantage. For example, Hepatitis C Virus requires a host miRNA expressed in the liver, mir-122, which interacts with the 5′-non-coding region of the viral genome to increase replication. Moreover, the host innate immune response downregulates mir-122, presumably reducing the viral replication.

31In the viRNA pathway, short RNAs are derived from the virus itself and will thus always match the target. This means that there is no possibility of the viral genome evolving to avoid sequence-based recognition. This is clearly important, as when a ‘fixed’ siRNA sequence is experimentally applied (i.e. one that is exogenous, rather than derived from the virus), the targeted viral sequence rapidly evolves to avoid being targeted. By contrast, in the miRNA pathway both host- and virus-encoded miRNAs may have an opportunity to coevolve with the target sequences in their opponent. As viruses generally evolve faster, thanks to their larger population sizes, shorter generation times and higher mutation rates, they may often have the upper hand.

32Host-encoded miRNAs involved in the control of gene expression are often conserved over long periods of evolutionary time in both plants and animals. Superficially, this might suggest that miRNA sequences are not engaged in rapid arms race evolution. However, recent sequencing of Drosophila miRNAs, in conjunction with analysis of 12 Drosophila genomes, suggests that some miRNAs are evolutionarily young and can be frequently gained and lost, opening up the possibility that there may be a class of rapidly evolving miRNAs that could target and coevolve with viruses.

33The VSRs of plant viruses typically suppress the miRNA pathway as well as the viRNA pathways, and viral suppression of the host miRNA functions can be a major cost of infection. Similarly, many viruses of vertebrates encode genes that suppress both RNAi pathways, possibly as a by-product of suppressing the interferon response. Given the possibility of antiviral miRNAs, it now seems plausible that this is an adaptive strategy. However, because many viruses use the miRNA pathway for their own advantage, a trade-off may occur between a virus' ability to suppress RNAi, and its ability to use miRNAs to its own advantage. Such a trade-off might not occur in Drosophila, where there is a separation of the core miRNA and viRNA pathways, and VSRs do not impede the miRNA pathway in vivo.

34TEs copy themselves throughout the genome, and can harm the host when they cause mutations or ectopic recombination (reviewed by ). RNAi can act as a defence against TEs, both by degrading transcripts, and by preventing expression by heterochromatin formation. Heterochromatin is usually strongly enriched for TEs, and although it is required for cellular functions such as the control of gene expression and successful chromosome segregation, it has been widely suggested that it originally evolved as an anti-TE defence.

35The de novo initiation of heterochromatic silencing, including that targeted against TEs, seems almost universally mediated by RNAi-related pathways, and in many organisms RNAi also plays a central role in its ongoing maintenance. In Arabidopsis, more than half the endogenous siRNAs are derived from heterochromatic sequence, and a third of these target TEs and other repeats. Here, the 24–26 nt siRNAs that target TEs are produced by a Dicer family member (DCL3), and then bound by an Ago-4-containing complex that targets the genomic copies of the TE for methylation and heterochromatin formation. A host RdRp (RDR2) is also required, suggesting that there may be an amplification step. Although it is not yet certain how this anti-TE RNAi is initiated, it has been suggested that dsRNA may be generated both by antisense TE transcripts expressed by host promoters, and by fold-back structures such as those resulting from inverted terminal repeats produced by some classes of TE.

36In vertebrates and Drosophila the piRNA pathway (figure 1c) appears to be a major regulator of TE activity in germ line-related cells. In Drosophila there are three Piwi-class Argonaute family proteins (Argonaute-3, Piwi and Aubergine) that are strongly expressed in the germ line. All three of these genes bind piRNAs that are primarily derived from TEs, and mutants show an elevated rate of transposition. Similarly, mutants in two of the three mouse Piwi homologues lead to increased expression of some TEs. Several more RNAi-related genes that are required for germ line TE silencing have been identified in the piRNA pathway of Drosophila (reviewed by ). The putative nucleases Zucchini and Squash are thought to be required for 3′-processing in the generation of mature piRNAs, and 3′-methylation by Hen1 stabilizes mature piRNAs. Additionally, mutations in Spindle-E/Homeless and Armitage (which are thought to be helicases), Krimper (a tudor domain protein), and Maelstrom all disrupt piRNA formation and increase the activity of retrotransposons.

37In C. elegans, several RNAi-pathway genes are necessary for silencing DNA transposons in the germ line, and numerous siRNAs are derived from TE sequences. However, although Piwi family Argonautes are found in C. elegans and are required for germ line function, with the exception of the TE Tc3, the C. elegans Piwi-related genes (PRG-1 and PRG-2) do not appear to be required for the suppression of TE transposition.

38Very recently, a second anti-TE RNAi pathway has been identified in Drosophila. This pathway uses Ago-2 and Dicer-2 from the viRNA pathway, and processes the TE transcripts into endogenous siRNAs in both somatic and germ line tissue. Although it has not yet been shown to affect rates of transposition, both Ago-2 and Dcr-2 knockouts display increased levels of TE transcript, strongly suggesting that this may be a second regulatory mechanism. Why two different RNAi pathways should be required to suppress TEs is unclear: it is possible that there is crosstalk between the pathways, and it has been hypothesized that the second pathway may be involved in the maintenance of heterochromatin in somatic tissue.

39In the germ line, the conflict between the host genome and TEs could drive an evolutionary arms race, in the same way as has been hypothesized for viruses. In support of this, a recent analysis of the genome sequences of 12 species of Drosophila found that natural selection has driven the rapid evolution of both Krimper and Spindle-E. To examine the other piRNA-pathway genes, we have summarized the rate of protein evolution in several components of the piRNA pathway (figure 3), and have tested these genes for effects of positive selection using the DNA sequence polymorphism data from the genome sequences of several D. simulans strains (figure 4). We found that Maelstrom, Krimper and Aubergine are all among the top 5 per cent of the most rapidly evolving genes in the genome, and Armitage, Aubergine, Piwi, Maelstrom and Spindle-E show evidence of positive selection (natural selection fixing advantageous amino acid substitutions; Krimper is not significant in this test, but does appear to have an extremely high level of amino acid polymorphism within D. simulans). This mirrors what is seen for the viRNA-pathway genes Ago-2, Dcr-2 and R2D2, and is in contrast to the miRNA-pathways genes that evolve very slowly (see §4).

40While the core machinery of the RNAi pathways (figure 1) is conserved in plants and animals, a fundamental difference between the kingdoms has emerged from studies of signal amplification and non-cell-autonomous RNAi. In particular, not all animal lineages carry key genes implicated in signal amplification and non-cell-autonomous RNAi (figure 5), and these differences are reflected in the diversity of RNAi phenotypes among animals (reviewed by ). Given the importance of signal amplification for antiviral RNAi in plants, this might be expected to have significant implications for the contribution RNAi is able to make to innate immunity in animals.

41The distribution of RdRps, which are thought to be required for signal amplification, is a striking example. With extensive genome data now available from all major metazoan groups, we have been able to identify this gene in lineages on each major branch of the bilateria, consistent with its presence in the metazoan ancestor (figure 5). Although the gene has been retained in some Porifera, the Cnidaria and some of the Protostomia (e.g. nematodes and the tick) and Deuterostomia (hemichordates, cephalochordate and a tunicate), in most of these groups, it has been independently lost from at least one representative. Among the lineages known to have lost the gene are the vertebrates and insects, as well as the others shown in figure 5. The RdRp genes found in all animal, fungal and plant genomes are characterized by a highly conserved domain (see figure 1 in the electronic supplementary material).

42SID-1, which was first identified in nematodes and forms a channel for cell-to-cell movement of dsRNA, has no orthologue in plants, but it does in many animal lineages—with the notable exception of the dipterans (figure 5). Compared to RdRp, the gene is not only more widespread, but the number of genes in each metazoan group is more variable. In the arthropods, for example, a single orthologue of SID-1 is found in each of the bee, louse and aphid genomes (figure 5), while multiple (three) orthologues are found in the genomes of the moth Bombyx and the beetle Tribolium. Furthermore, the relationships of the insect SIDs do not match the phylogeny of the animals they are found in. The bee SID-1 protein clusters with one of three beetle and two of three moth SID homologues (fig. 12 in ), as well as the two vertebrate SID paralogues, one of which was lost in fish, but preserved in birds and mammals. This suggests that there were ancient duplications of this gene, and some of the resulting copies have subsequently been lost in some lineages. The variation in the number of SID-1 genes may explain why the effects of RNAi seem so variable in different species. For example, if RNAi is used experimentally to knock down gene expression, it can result in systemic silencing in moths and beetles, while in flies, which have no SID-1-like genes, transgene silencing has only a localized effect. A more recent study using multiple RNAi of all SID-1 orthologues in Tribolium has cast a doubt on whether these genes function in systemic RNAi at all. The basis for the varying, albeit limited, extent of systemic RNAi in arthropods therefore remains unexplained. Understanding the function of the different arthropod SID genes will be of great interest, as even in C. elegans the function of only two of the several SID genes is known.

43Dicer genes also show considerable variation among eukaryotic lineages. In plants, for example, there are larger numbers—Arabidopsis has four Dicer-like genes—and these show some functional redundancy (reviewed by ). For example, resistance phenotypes against some viral infections are seen in Arabidopsis only when multiple Dicers are knocked out, suggesting that they can substitute for each other. In Arabidopsis, it appears that the Dicer DCL4 is a primary effector against ssRNA viruses, but that DCL2 substitutes when DCL4 is knocked out or suppressed, and DCL3 (which normally mediates chromatin modification) plays an important role in some viral infections, and in DCL2/DCL4 double mutants. This may suggest that, in plants at least, multiple paralogous viRNA-pathway genes have specialized functions, perhaps evolving in response to viral suppression of RNAi. In animals, one or two Dicers have been found in all genomes explored to date, with the exception of Daphnia (figure 5; ). The vertebrates and nematodes have a single Dicer responsible for both siRNA and miRNA production. Genes for two Dicers have been found in first Drosophila, and now in the honeybee and Tribolium genomes.

44Even more striking is the variability in the components of the RISC complexes. Although it is not always clear that they function within RNAi-related pathways, Argonaute family members, in particular, seem to be very variable in number—Arabidopsis has 10, fission yeast (Schizosaccharomyces pombe) 1, Drosophila 5, and C. elegans in excess of 20. Why this should be is not clear. Although there is sometimes pathway specialization (Piwi-like Argonautes function in the piRNA pathway), there is also overlap between the pathways, and genes can be functionally redundant. In the nematode, there is discrimination in loading siRNAs or miRNAs onto RISCs after processing by its single Dicer, so that short RNAs with different functional roles are associated with different members of the large family of Argonautes. The siRNAs are found bound to RDE-1, which has RNase H activity; by contrast, miRNAs are bound to the Argonaute proteins ALG-1 and ALG-2. Insects and vertebrates appear to show less discrimination in this respect. In Drosophila, miRNAs and siRNAs are generated by distinct Dicers but the duplex intermediates undergo independent sorting processes, one involving Dcr-2/R2D2 that favours loading of siRNAs into Ago-2, and may be loaded into Ago-1 or into Ago-2. Ago-1 has inefficient RNase H activity, so that only Ago-2 is capable of cleaving target RNAs, where there is sufficient complementarity. In humans, which also have only one Dicer, miRNAs may be loaded onto any of the four proteins Ago-1–4; only Ago-2 has RNase H activity.

45Despite evolution and arms races being frequently mentioned in the RNAi literature, there are still few explicitly evolutionary studies of RNAi. In this final section, we highlight three interesting topics for future work.

46In this review, we have discussed how coevolution between hosts and parasites can take the form of an arms race, in which novel mutations that either increase resistance or overcome this resistance sweep through the host and parasite populations. However, coevolution can also act to maintain genetic variation in the resistance of individuals in host populations and infectivity of strains in populations. This occurs because as the frequency of a resistance allele increases in the host population, the frequency of the pathogens it targets and hence its selective advantage can decrease. If the resistant allele confers susceptibility to other pathogen genotypes, or has some other fitness cost, there comes a point where it is at a selective disadvantage. An analogous process can occur within the parasite population. This results in negative frequency-dependent selection, and can maintain genetic variation in populations and promote complex dynamics, such as cycles in allele frequency.

47It is possible that this form of coevolution could maintain variation in RNAi genes and viral populations. For example, because host-encoded miRNAs that target viral sequences are a form of specific immunity that is germ line encoded (as opposed to viRNAs, where immunity is acquired after infection), antiviral miRNAs will select for escape mutants in the viral population, and these escape mutants could in turn select for matching miRNAs. This could lead to specific ‘matching alleles’ in the host and viral population, in which each host allele confers resistance to some viral genotypes but susceptibility to others, resulting in the coevolutionary dynamics described above This could both maintain genetic variation in miRNAs and their viral targets, and lead to dynamic changes in the effectiveness of miRNA defences through time and space.

48Unlike the miRNA pathway, where immunity is germ line encoded and may take many generations to evolve, highly specific viRNA-based immunity can develop rapidly after exposure to the virus. In this respect, the viRNAi pathway is reminiscent of the acquired immune response of vertebrates, which is also highly specific and develops after infection. In vertebrates, acquired immunity can maintain a high diversity of antigens in the parasite population—hosts are most likely to be immune to common antigenic types, so there is an advantage to rare antigen alleles. A similar process can occur within an infected host, where mutations in antigens that escape the immune response may be favoured, leading to a turnover of antigens through time. Could such processes occur in response to the viRNAi pathway? Plants that have been infected by a virus can develop new growth that is free from viral infection and resistant to secondary infection due to RNAi targeted against the virus. If these hosts are exposed to a second viral strain, then this incoming strain may have an advantage if it is not recognized by viRNAs derived from the resident strain. This could potentially maintain genetic polymorphisms in the viral population. Similarly, there may be selection for mutants that escape recognition by the viRNA pathway within an infected host, by evolving new (and thus rare) sequences. However, the acquired immune response of vertebrates and viRNA differ in the speed of the response, whether or not it is systemic, and the number of different epitopes or viRNAs produced from a pathogen, and these factors may all affect the outcome. These ideas could be tested both theoretically and experimentally.

49In Drosophila, antiviral and anti-TE RNAi genes are among the most rapidly evolving genes in the genome (figures 4 and 5), and both viruses and TEs select for resistance in their hosts. However, while it is easy to see how the rapid evolution of antiviral genes might be driven by VSRs, it is less clear how TEs could impose such strong selection, as they are not known to suppress RNAi (although this possibility has been proposed, see ).

50An alternative explanation for rapid evolution in the Drosophila piRNA pathway might be that some piRNA components also have antiviral function in addition to their anti-TE role. In support of this, some piRNA-pathway mutants appear to be more susceptible to a dsRNA virus (although these results await further experimental confirmation; ), and there is an extensive overlap between the antiviral and anti-TE pathways in other organisms (e.g.; ). Moreover, components of the piRNA pathway might also interact with DNA viruses and retroviruses. First, DNA virus function often depends upon chromatin status (e.g. ) and in plants this can be mediated by RNAi. Second, the distinction between retroviruses and retrotransposons is a subtle one (retrotransposons are often considered ‘endogenous retroviruses’; e.g.; ) and in Drosophila retrotransposons are silenced by the piRNA pathway in the germ line, opening up the possibility that the piRNA pathway could also suppress ‘true’ retroviruses.

51It is also possible that a factor other than viruses or TEs drives rapid adaptive evolution in both RNAi pathways. For example, both Ago-2 and Dcr-2 have other germ line functions in Drosophila, and the Piwi family genes of both C. elegans and Drosophila have non-TE-related germ line functions, such as mediating the silencing of other potentially ‘selfish’ sequences such as Stellate in Drosophila. Indeed, in general, the most rapidly evolving genes in Drosophila are those that regulate chromatin, mediate nuclear import and export, and are involved in meiosis and reproduction, and it may be that rapidly evolving RNAi genes are a part of this phenomenon. Thus, although viruses remain a prime candidate, it remains an open question as to whether viruses, TEs, other parasites, meiotic drive, sexual conflict or something else entirely has driven the adaptive evolution seen in RNAi-pathway genes.