The first microRNAs (miRNAs) to be discovered, lin-4 and let-7, were found to be regulators of Caenorhabditis elegans development 1 2 3 , and they established a paradigm for eukaryotic gene regulation in which short hairpins generate RNAs of approximately 22 nucleotides (nt) that repress specific target mRNAs. miRNAs have proved to be pervasive in both animals 4 5 6 and plants 7 8 , acting as sequence-specific guides for target recognition 9 10 . Several thousand miRNAs have now been found in dozens of plants and animals 11 . Moreover, the biogenesis and activity of miRNAs are strongly related to those of small interfering RNAs (siRNAs) that mediate RNA interference, another ancient mechanism for post-transcriptional gene silencing 12 .

Although miRNAs mediate diverse aspects of development and physiology in both plants and animals 13 14 , there are substantial differences between them. For example, the loci that produce miRNAs have distinct genomic arrangements in each kingdom. Furthermore, miRNAs are excised from precursor transcripts by different pathways in the two kingdoms, and in different subcellular compartments. Once made, plant and animal miRNAs have vastly different suites of direct targets; the number of direct targets of a given animal miRNA generally exceeds that of a given plant miRNA by at least an order of magnitude 15 . Herein, we focus on how these differences contribute to, and are the result of, distinct evolutionary characteristics of miRNAs in the two kingdoms. We also highlight many commonalities between the respective systems that may reflect a shared evolutionary heritage or convergent strategies for handling and metabolizing double-stranded RNAs.

What is a miRNA? Answering this question is not a simple task, as no single definition clearly and specifically encompasses all miRNAs. Although practical guides for miRNA annotation in plants and animals exist 16 17 , not all loci reported in the miRBase registry 18 have been annotated to the same degree of confidence. In general, miRNAs are the products of inverted repeat transcripts that are precisely cleaved by RNase III enzyme(s) in the Dicer and/or Drosha protein families to yield small RNAs of approximately 21 to 24 nucleotides that guide Argonaute (AGO) proteins to complementary targets. Analogous, but distinct, core pathways govern the biogenesis of most miRNAs in plants and animals. Although we focus on these canonical miRNA pathways, a plethora of alternative pathways exist. Indeed, the diversity and flexibility of miRNA biogenesis pathways, in concert with related mechanisms that generate siRNAs, have made a significant contribution to miRNA evolution. In addition, while a hallmark of most studied miRNAs is the precise manner in which they are excised from precursor hairpins, there are examples of imprecisely cleaved miRNAs. As we shall see, this phenomenon might have implications for miRNA evolution, but it also poses challenges for the accurate distinction of bona fide miRNAs from fortuitous hairpins associated with short RNAs not generated by a specific biogenesis machinery.

The substantial differences between the biogenesis of animal and plant miRNAs are also reflected in the differences in their requirements for target recognition. It has long been known that plant miRNAs often have targets with perfect 72 or, more frequently, near-perfect 10 complementarity, facilitating relatively simple identification. Canonical plant miRNA target sites are found in 5' UTRs, ORFs and 3' UTRs, as well as within non-protein-coding transcripts, suggesting that all RNA contexts are equally amenable to miRNA-directed regulation in plants. Many of these plant miRNA targets succumb to AGO-catalyzed cleavage when they encounter a cognate miRNA; the characteristic remnants of these cleavage reactions enable molecular confirmation of plant miRNA target predictions in vivo 72 73 74 . However, not all plant miRNA-target interactions lead to AGO-catalyzed slicing. Some plant miRNA targets have conserved central mismatches embedded within perfectly base-paired regions at the 5' and 3' ends that allow AGO/miRNA binding but prevent slicing 75 76 , whereas others seem to be refractory to target cleavage despite extensive complementarity 77 . In addition, even for plant miRNA targets that are sliced, slicing is often not the sole mechanism by which miRNAs repress target gene expression in plants: several experiments, involving multiple plant miRNA families, have demonstrated a pervasive contribution of translational repression to plant miRNA function 77 78 79 .

Whether operating via slicing and/or translational repression, though, most evidence indicates that plant miRNAs require extensive pairing to their targets. By contrast, it is rare for animal miRNAs to identify targets with 'plant-like' complementarity. The initial target sites identified for the founding miRNA lin-4, within the 3' UTR of C. elegans lin-14, exhibited only partial complementarity 1 2 . Subsequent studies in the Drosophila system elucidated arrays of approximately seven-nucleotide conserved 3' UTR motifs termed Brd boxes, GY boxes and K boxes, which mediate critical post-transcriptional repression of Notch target genes during sensory bristle and eye development 80 81 82 . These motifs proved to represent binding sites for most of the initially described Drosophila miRNAs 5 , and defined their capacity to identify targets via complementarity to their 5' ends, preferentially at nucleotides 2 to 8 9 . Extensive computational and experimental studies verified this as the major mode of miRNA target recognition in animals, with Watson-Crick pairing of positions 2 to 8 of the miRNA referred to as 'seed-pairing' 83 84 . Additional features, such as an adenosine following the seed match, the location within the 3' UTR, proximity to other miRNA-binding sites, and the degree of local secondary structure, also influence target site activity 15 .

Although there is also clear evidence for evolutionary selection of animal miRNA-binding sites in coding regions or even 5' UTR sites 85 86 , most of the well-studied target sites in this kingdom occur in 3' UTRs. The efficacy of CDS or 5' UTR sites appears to be hampered owing to competition with ribosomes 87 . Since artificial siRNAs that guide target slicing operate efficiently via coding sites in animal transcripts, similar to plant miRNAs, the paucity of CDS targeting is consistent with the view that relatively little miRNA target regulation in animals is mediated by AGO-catalyzed slicing; indeed, efforts to identify sliced remnants of AGO-catalyzed cleavage in mammalian samples yield very few targets 88 89 . The molecular mechanism(s) of miRNA targeting remain under investigation, and many have been proposed 90 91 . Although much attention has been focused on translational inhibition mechanisms, there is also evidence that bulk regulatory properties of animal miRNAs can be explained by mRNA degradation, possibly through induction of deadenylation 80 81 92 93 .

In vertebrates, about 30% of transcripts contain probable miRNA-binding sites that have been conserved between mammals and chicken 94 95 , and a comparable breadth of targeting has been detected in invertebrates 34 96 97 . Genome-wide transcriptome 92 98 and proteome 99 100 studies provide experimental support for the breadth of miRNA targeting in animal cells, and further indicate that many functional miRNA:target interactions are not well conserved. Moreover, at least some functional animal miRNA binding sites lack seed-pairing 2 3 101 . miRNA targeting in animals has broad potential to be combinatorial, since individual targets often bear conserved target sites for different miRNAs 94 95 . The scope of miRNA targeting appears to be drastically different in plants. Less than 1% of the transcripts in Arabidopsis are known or predicted miRNA targets 20 74 and there appears to be little, if any, combinatorial control. Almost all known plant miRNA targets have a single target site and are regulated by just one miRNA. These genome-wide principles may support the notion that animal miRNAs generally cast a wide net of mostly subtle regulatory effects across the transcriptome, while plant miRNAs have more focused and stronger regulatory effects on a relative handful of key targets. Of course, these generalizations should not be over-interpreted. Although quantitatively mild regulation can be of substantial importance to normal development or physiology, the loss of potent regulatory interactions is sometimes of surprisingly minimal phenotypic consequence 13 .

Overall, the pairing requirements and target breadth of animal and plant miRNAs are clearly distinct, and this might be related to differences in their evolutionary emergence. Nevertheless, there is reason to believe that evolution has acted upon a backdrop of shared ancestral mechanisms. A recent study showed that Drosophila endo-siRNAs loaded in AGO2 require their 2'-O-methylation to remain stable in the presence of highly complementary targets 102 . Reciprocally, instability of Drosophila miRNAs loaded in AGO1 was induced by providing them an artificial, perfectly complementary target, and similar findings applied to mammalian miRNAs. It was proposed that the general rarity of highly complementary targets for animal miRNAs has permitted them to dispense with a 3' protection pathway 102 . By contrast, the fact that most plant miRNAs do have highly complementary targets may provide a pressure for obligate 3' methylation of plant miRNAs. Moreover, the piwi-interacting RNA (piRNA) class of small RNAs carried by animal Piwi proteins are also methyated by Hen1 42 103 104 , and key functions of piRNAs include the recognition of perfectly matching transposon transcripts 105 . These data suggest that an evolutionarily ancient aspect of small regulatory RNA-mediated regulation is sensitive to the status of target pairing.

The genomes of most higher eukaryotes are predicted to encode at least 10⁵ to 10⁶ putative hairpins with substantial similarity to validated pri-miRNA hairpins 122 123 124 . In theory, this constitutes an enormous reservoir of putative miRNA substrates, whose trace regulatory activities might be subject to selection and evolutionary stabilization. But how many predicted hairpins in a given genome are actually competent to be specifically processed by the miRNA biogenesis machinery? Conserved miRNA genes are amenable to computational discovery by signatures of hairpins that exhibit evolutionarily stable arms and diverging terminal loops 122 123 . However, it is not currently possible to prospectively annotate the miRNAs encoded by a genome, in the absence of comparative genomics, with any reasonable degree of specificity or sensitivity. Therefore, the mere existence of large numbers of predicted miRNA-like hairpins does not imply the existence of similar numbers of evolutionarily emergent miRNA genes.

In light of broad transcription across euchromatin of plants and animals 125 126 , including substantial amounts of transcribed sequence that are removed by quality control mechanisms, it seems inevitable that short pieces of RNA will eventually be associated with most of the genome, including most predicted genomic hairpins. For example, many hundreds of candidate miRNA hairpins in various Drosophila species were initially annotated on the basis of singleton reads 119 . However, these were either not recovered in substantially larger datasets from the same tissue sources, or generated heterogeneously sized reads mapping through the predicted hairpins 127 , indicating that few of these were genuine miRNAs produced by RNase III processing.

More recently, short RNAs were recovered from >100,000 hairpins in the D. melanogaster genome (from a starting set of nearly 1 billion short RNA reads from almost 200 libraries) 52 . However, confident canonical miRNA production could only be assigned to approximately 200 loci. These included many miRNA hairpins that are recently or newly evolved in D. melanogaster, but they did not include hundreds of other miRNA candidates predicted from comparative analysis of the 12 sequenced Drosophilids 34 128 . Even though most of these predictions were associated with at least some small RNA reads, none exhibited read patterns and/or sizes that were consistent with Drosha/Dicer-mediated processing (Figure 3). Although caution should be exercised in interpreting negative evidence, such data support the notion that a fairly limited number of genomic regions are competent for recognition by the miRNA pathway when transcribed.

These findings are echoed by studies of other animals and plants. For example, while novel mammalian miRNAs continued to be reported, a recent systematic analysis actually revised the estimates of mammalian miRNAs downward, and called into question a substantial subset of extant annotations 129 . Similarly in Arabidopsis, many hundreds of miRNA genes were predicted by comparing hairpins with inferred target sites 130 . However, very few of these have since been validated by deep sequencing 20 45 . Altogether, these observations suggest that relatively few of the many predicted genomic hairpins are substrates of miRNA biogenesis pathways, and hint that substantial bioinformatic progress remains to bring miRNA prediction up to a par with protein-coding gene prediction.

It is clear that more data on the efficiency and specificity of miRNA processing are desirable, especially as all entries in the miRBase registry are currently treated as equivalent 11 . Although some loci are now deemed suspect, in principle, the majority of entries are confident miRNA genes that exhibit reasonably precise processing. The reality is that with ever-increasing depths of small RNA sequencing, all miRNA loci exhibit some level of terminal heterogeneity in their cloned products, and this tends to blur the division between loci that can be confidently inferred to have transited an RNase III pathway acting on a precursor hairpin, as opposed to fortuitous hairpins that generate degradation products (Figure 3a). At present, a cautious approach to miRNA annotation seems warranted, in which miRNA 'candidates' are segregated from more confident miRNA loci 52 . These include loci that exhibit plausible miRNA/miRNA* duplexes but have substantial reads not conforming to RNase III products, as well as hairpins lacking a star species (Figure 3b). In particular, the lack of cloned miRNA* species was not previously seen as an impediment to miRNA annotation; however, the depth of next-generation sequencing now makes it reasonable to demand miRNA* species for confident annotation. (Exceptions may be made with appropriate data; that is, if there is strong evidence from loss of reads in miRNA biogenesis mutants, or enrichment of reads in Ago-IP samples.)

Perhaps more importantly, there are few data on how 'well' a miRNA hairpin is processed. If only 1% of a miRNA hairpin transits the processing machinery, and these are processed precisely, then it is equally eligible for annotation as loci for which 100% of the hairpin is converted into miRNAs. The efficiency of processing cannot be inferred from counts of small RNAs alone, since some endogenous miRNA loci might generate rare reads owing to cell-specific expression, or perhaps post-transcriptional inhibition of some aspect of their biogenesis 43 . It is only recently that efforts have been made to generate systematic data on the effectiveness of miRNA processing, by cloning small RNA libraries from pools of cells transfected with different miRNA expression constructs and quantifying the output reads 129 . Larger scale data on forced expression of predicted hairpins, especially ones that lack endogenous reads or exhibit heterogenous reads, should provide valuable insights into the potentially partial capacity of some substrates to enter miRNA biogenesis pathways.

In general, small RNAs that populate a miRNA-effector AGO have the functional attributes of a miRNA, regardless of whether it was produced by a canonical or alternative pathway. However, this does not mean that all miRNA substrates evolve similarly. For example, the evolution of plant miRNAs from target duplications can be compared with those from incidental hairpins. Because the plant miRNA system appears to require extensive target complementarity, it is presumed that only rarely will fortuitous small RNA-generating hairpins emerge and then subsequently acquire beneficial targets. Therefore, these classes of plant miRNAs are expected to emerge and disappear with distinct dynamics 106 107 120 121 .

Does the genomic location of miRNA hairpins influence their emergence? Canonical miRNAs in animals, unlike in plants, are commonly located in introns. Because animal miRNAs mostly emerge from incidental foldbacks, the introns of bona fide transcription units might be a privileged location for miRNA birth already endowed with directed primary transcription. We have also mentioned that animal miRNAs, unlike their plant brethren, are frequently arranged in operons composed of dissimilar species. This suggests that the emergence of animal miRNAs is privileged by location near an extant miRNA (Figure 2c). Access to the Drosha/DGCR8 processing complex is the gatekeeper for entry into the canonical miRNA pathway, and its activity has been reported to be co-transcriptional 131 132 133 . Consequently, physical proximity of an emergent hairpin to an established miRNA hairpin could enhance its nuclear cleavage relative to 'solo' emergent hairpins; this might be particularly important for the processing of suboptimal hairpins. Newly evolved miRNAs have indeed been detected in proximity to much more deeply conserved animal miRNAs 52 . The lack of Drosha homologs in plants might render this characteristic less relevant in the plant kingdom.

What is the influence of the biogenesis mechanism on miRNA evolution? If we compare the canonical miRNAs with mirtrons (Figure 2d), a subset of mirtrons are well conserved in a given animal clade (that is, Drosophila or vertebrates), and these diverge more quickly in their terminal loop than their hairpin arms 55 134 , as seen with canonical miRNAs. However, relatively few mirtrons are well conserved and no mirtrons are common between invertebrates and vertebrates, whereas many canonical miRNAs are identical over this distance. This implied that mirtrons, as a class, emerge and disappear more quickly than canonical miRNAs. This was confirmed with deep sequencing of small RNAs from three Drosophilid species; this showed that mirtrons comprise a steadily increasing fraction of confidently annotated miRNA loci as the evolutionary branch length under consideration decreases 127 . Recent analysis of mirtrons from ultradeep sequencing has extended the species-specific catalog of fly and nematode mirtrons further still 52 135 .

The distinct evolution of mirtrons and canonical miRNAs might relate to the structural hurdles needed to become a substrate of the respective biogenesis machineries (Figure 2d). In the case of canonical miRNAs, a substrate must simultaneously adopt conformations that permit its cleavage by both Drosha and Dicer. In the case of mirtrons, a substrate must be spliced and be a target of Dicer. We do not know how many endogenous substrates of Drosha there are, but there are currently no more than some hundreds known in any given animal species. By contrast, the many short constitutively spliced introns (for example, there are 27,000 to 30,000 introns <120 nucleotides in flies and nematodes) comprise a large pool of loci that have passed one processing hurdle. Structure-function studies have shown that mirtrons are flexible with regard to primary sequence, provided that they retain splicing functionality and adopt substantial hairpin structure with 3' overhangs 135 . In fact, the nature of splice sites (GU...YAG) should position many introns to pair the 5' G with the 3' Y, leaving the AG as a two-nucleotide 3' overhang.

These factors might conspire to aid the evolutionary emergence of animal mirtrons, relative to canonical miRNAs. Conversely, the apparent near absence of mirtrons in plants could well be a consequence of a distinct biogenesis mechanism that does not rely upon processing by two separate RNase III enzymes. Because plant miRNA maturation relies upon a single Dicer-like protein to completely liberate a miRNA/miRNA* duplex from a pri-miRNA transcript, the pre-miRNA intermediates produced by mirtrons may be either unnecessary or perhaps even unrecognized by the plant DCL1 complex.

The major plant and animal miRNA pathways differ with respect to their biochemical mechanisms, the extent of their preferred target pairing, and numbers of functional targets. These differences have resulted in distinct characteristics of the evolution of plant and animal miRNAs. In particular, the co-evolution of target:miRNA pairs is common in plants, whereas it seems much more common for animal miRNAs to emerge and then acquire target genes. One interpretation is that this reflects independent emergence of miRNA pathways in plants and animals, on the backbone of an ancestral RNAi pathway that metabolized dsRNA into short RNAs that populate Argonaute proteins. This system might have emerged to defend against invasive nucleic acids such as viruses and transposons, and subsequently been adapted to generate miRNAs from endogenous inverted repeat transcripts. However, there are also analogies between plant and animal miRNA pathways. For example, certain vertebrate miRNA targets, as well as Drosophila hpRNA targets, exhibit 'plant-like' extensive complementarity. There is reciprocally a growing appreciation that plant miRNAs have emerged from incidentally emerged hairpins, akin to the presumed dominant mode for animal miRNA birth. Therefore, an alternative interpretation is that a miRNA pathway was extant in the last common ancestor of plants and animals, but became differentially deployed in these kingdoms.

In either case, it is clear that a limited set of core proteins, namely RNase III enzymes and Argonaute proteins, have been joined in remarkably diverse ways to control gene expression via small RNAs. Recent studies of fungal small RNA pathways provide additional evidence for innovation of RNase III-independent mechanisms for siRNA and miRNA production 136 137 , for which we can only guess at the underlying reasons that permitted the loss of canonical pathways and invention of new pathways. Altogether, it is evident that miRNAs are not a unitary entity, but instead encompass a variety of conceptually related phenomena, whose evolutionary pressures differ according to mechanism of biogenesis and even genomic location. Understanding the principles that govern the evolutionary flux of these myriad small RNA pathways will provide a fundamental complement to understanding the flux of protein-coding genes 138 .

bp: base pair; CDS: coding sequence; endo-siRNa: endogenous siRNA; hpRNA: hairpin RNA; miRNA: microRNA; MITE: miniature inverted-repeat transposable element; ORF: open reading frame; piRNA: piwi-interacting RNA; pre-miRNA: precursor miRNA; pri-miRNA: primary miRNA; siRNA: small interfering RNA; UTR: untranslated region.