Regulation of microRNA function in animals

Abstract

Since their serendipitous discovery in nematodes, microRNAs (miRNAs) have emerged as key regulators of biological processes in animals. These small RNAs form complex regulatory networks in cell development, differentiation and homeostasis. Deregulation of miRNA function is associated with an increasing number of human diseases, particularly cancer. Recent discoveries have expanded our understanding of how miRNAs are regulated. Here we review the mechanisms that modulate miRNA activity, their stability and their localization through alternative processing, sequence editing, post-translational modifications of Argonaute proteins, viral factors, transport from the cytoplasm and regulation of miRNA–target interactions. We conclude by discussing intriguing open questions to be answered by future research.

Introduction

microRNAs (miRNAs) are short non-coding RNAs of ~22 nucleotides that mediate gene silencing by guiding Argonaute (AGO) proteins to target sites in the 3’ untranslated region (UTR) of mRNAs. AGOs constitute a large family of proteins that use single-stranded small nucleic acids as guides to complementary sequences in RNA or DNA targeted for silencing¹. The miRNA-loaded AGO forms the targeting module of the miRNA-induced silencing complex (miRISC), which promotes translation repression and degradation of targeted mRNAs². The first miRNA was discovered in Caenorhabditis elegans as a short RNA produced of the lin-4 gene, which post-transcriptionally represses the lin-14 mRNA^3,4. Such small RNAs were widely thought to be unique to nematodes, until they were shown to be abundant in diverse animal phyla⁵. This new class of regulators was subsequently named ‘microRNAs’^6–8.

miRNAs are involved in virtually every cellular process and are essential for animal development, cell differentiation and homeostasis; deletions of the fundamental miRNA biogenesis factors Dicer⁹ and Drosha¹⁰ are lethal in mouse embryos. Although the importance of miRNAs for embryonic development is well established¹¹, Drosha and Dicer are involved in other nuclear processes, such as pre-mRNA splicing, which may also contribute to their deletion phenotypes¹². The miRNA repository miRbase (www.mirbase.org) currently lists 1,917 precursor miRNAs (pre-miRNAs) and 2,654 mature miRNAs in Homo sapiens¹³, and more than 60% of human protein-coding genes harbor predicted miRNA target sites¹⁴. Deregulation of miRNA function is associated with numerous diseases¹⁵, particularly cancer^16,17: miRNAs can be both oncogenes (called oncomirs) ¹⁸ or tumor suppressors¹⁹, although overall down regulation of miRNA expression is a hallmark of cancer²⁰. Some miRNAs are prognostic markers^21,22 or potential targets for novel cancer therapies²³. Plant miRNAs^24,25, which are not discussed here, differ considerably from animal miRNAs in their evolution²⁶, biogenesis²⁷ and function²⁸(reviewed in ^29,30).

In this Review, we first provide an overview of miRNA function and regulation. We than discuss in detail the regulation of miRNA function through the formation of isomirs[G, glossary term] Where you agree please provide succinct, one-sentence definitions for these specialist terms.], the addition of non-templated nucleotides, post-translational modification (PTM) of miRNA-binding proteins, miRNA sequestration, modulation by viral factors, transport from the cytoplasm, and the regulation of interactions between miRNAs and their target mRNAs (miRNA–target interactions).

Overview of miRNA function

miRNA genes are transcribed into primary miRNA (pri-miRNA) transcripts and undergo multi-step biogenesis, in which they are processed first into pre-miRNAs and finally into mature miRNAs (BOX 1). miRNAs exhibit tissue-specific expression patterns³¹, which are primarily regulated transcriptionally³². They are transcribed mainly by RNA polymerase II^33–35 and can be derived from introns or from long non-coding RNAs. Pri-miRNAs can consist of a single mature miRNA or of clusters of often related miRNAs^6,36. miRNAs are grouped into families³⁷ based on the similarity of their seed sequences; the seed comprises nucleotides 2–8 (counting from the 5’ end) and is primarily responsible for miRNA targeting of mRNAs³⁸.

Box 1: miRNA biogenesis.

The biogenesis of microRNAs (miRNAs) is a multi-step process. They are transcribed mainly by RNA polymerase II ^33–35 as structured primary miRNAs (pri-miRNAs) and processed into precursor miRNAs (pre-miRNAs) and finally into mature miRNA duplexes (see the figure). The mature miRNA comprises a 5p strand (arising from the 5’ arm of the pre-miRNA hairpin) and a 3p strand. The sequence of pri-miRNAs can be altered by A-to-I editing by double-stranded RNA-specific adenosine deaminase (ADAR) proteins, which may affect further biogenesis and the sequence of the mature miRNA, or promote degradation of the pri-miRNA⁹⁹.

The pri-miRNA hairpin is excised in the nucleus by the Microprocessor complex comprising the RNase III enzyme Drosha²⁵² and the protein DiGeorge syndrome critical region gene 8 (DGCR8) ^253,254. Drosha recognizes the double-strand RNA–single strand RNA junction at the hairpin base, whereas two DGCR8 proteins bind the stem and ensure correct cleavage^255,256. Alternative cleavage by Drosha leads to the production of isomirs^97,98 (FIGURE 2; Supplementary Table S1).

Pre-miRNAs are hairpins of ~70 nucleotides^6–8. The hairpin end features a 2-nucleotide overhang at the 3’, a 5’ phosphate and a 3’ hydroxyl, which are typical of RNase III products²⁵⁷. Exportin-5 recognizes the overhang²⁵⁸ and transports the pre-miRNA to the cytoplasm^259–261. XPO5 knockout in a human cell line reduced but did not eliminate nuclear export of pre-miRNAs, which suggested that alternative modes of pre-miRNA nuclear export exist²⁶².

In the cytoplasm, the RNase III enzyme Dicer^259–261 binds the pre-miRNA by recognizing the 5’ phosphate, 3’ overhang and loop structure^263–266. Dicer is a ‘molecular ruler’^263,267 that cleaves pre-miRNAs at a species-specific length²⁶⁸ and yields a mature-miRNA duplex with another typical 2-nucleotide 3’ overhang^257,269. Alternative cleavage by Dicer can also lead to the production of isomirs⁹⁸ (FIGURE 2; Supplementary Table S1). In vertebrates, cleavage by Dicer is modulated by TAR RNA-binding protein (TARBP) and protein activator of the interferon-induced protein kinase (PACT); in flies, by Loquacious proteins^{103–106,270}.

One strand of the mature miRNA (the ‘guide’ strand) is loaded into Ago, whereas the other strand (‘passenger strand’) is discarded^271–273. Loading preference is given to the strand possessing the less stably paired 5’ end^274–276; Ago2 was also reported to prefer an A or U as the 5’-terminal nucleotide²⁷⁷.

The mature miRNA forms a functional unit with an AGO protein (FIGURE 1A), which is a single polypeptide chain composed of four characteristic domains: the amino-terminal (N) domain, the Piwi-Argonaute-Zwille (PAZ) domain, the middle (MID) domain and the P-element induced wimpy testes (PIWI) domain. Two linker domains (L1 and L2) connect the N and PAZ domains (L1) and the PAZ and MID domains (L2). The N and PAZ domains form one lobe of AGO and the MID and PIWI domains form the second lobe. The MID and PIWI domains hold the 5’ end of the miRNA, whereas the PAZ domain binds its 3’ nucleotide^39–42. The mammalian genome encodes four AGO proteins (Ago1–4). Ago2 is the most highly expressed and the only AGO protein able to cleave a target that is fully complementary to the guide strand[G] of the miRNA^43,44, which is a feature important for the biogenesis of miR-451^45,46 and for regulation of a subset of miRNAs that are extensively paired with their targets^47,48. Although some human miRNAs are preferentially loaded into specific AGO proteins, many associate with all AGO proteins^49–52.

Figure 1: Overview of miRNA function and its regulation.

(A) Mature miRNAs operate as functional units that include an Argonaute protein. Argonaute proteins have four domains, the amino-terminal domain (N), the Piwi-Argonaute-Zwille (PAZ) domain, the middle (MID) domain and the P-element induced wimpy testes (PIWI) domain, and two linker regions (L1 and L2). The MID and PIWI domain hold the 5’ end of the miRNA and arrange it in a helical conformation; the MID domain has a binding pocket for the 5’-terminal nucleotide. Another binding pocket, in the PAZ domain, holds the 3’-terminal nucleotide ^39–41. Nucleotides 2–8 from the 5’ end of the miRNA form the seed, which is crucial for target-mRNA recognition³⁸. Seed interactions involve nucleotides 2–8, 2–7 and 2–6 ⁵⁴, and can be supplemented by the binding to the MID domain of an adenine in the target mRNA opposite miRNA nucleotide 1 (t1) ^54,56,57, or through additional, base-pairing to nucleotides ~13–16 of the miRNA, which are termed ‘supplemental region’⁶⁵.

(B) miRNAs silence gene expression by inhibiting translation at the initiation step, likely through release of eukaryotic initiation factor 4A-1 (eIF4A1) and eIF4A2^79–81, and by mediating mRNA decay² through interactions with glycine-tryptophan protein of 182 kDa (GW182) proteins^68–71. GW182 binds polyadenylate-binding protein (PABPC) and the deadenylation complexes poly(A) nuclease 2 (PAN2)–PAN3 and carbon catabolite repressor protein 4 (CCR4)–NOT^72–76. Deadenylation is followed by decapping by the complex mRNA-decapping enzyme subunit 1 (DCP1)–DCP2⁷³ and 5’−3’ mRNA degradation (not shown) ⁷⁷.

(C) miRNAs form complex networks of interactions, as one miRNA can target many different mRNAs, ⁸⁶ and one mRNA can be regulated by many different miRNAs⁸⁷, with cooperative repression achieved by binding closely-spaced target sites^65,89,90.

miRNA target sites are generally located in the 3’ UTR of mRNAs; they possess strong complementarity to the seed region³⁸, which is the main criterion for target-site prediction^53–55. The strongest canonical (seed-matching) target sites are those that complement miRNA nucleotides 2–8 and have an adenine opposite miRNA nucleotide 1 (known as ‘t1A’), followed by those complementing nucleotides 2–8 without a t1A and nucleotides 2–7 with a t1A⁵⁴. t1A is not recognized by the miRNA guide strand, but by a binding pocket within AGO^56,57 (FIGURE 1A). Target sites with complementarity to nucleotides 2–7 or 3–8 of the miRNA are much weaker, but still considered canonical⁵⁴. Structural and single-molecule studies suggest that target recognition is achieved by a two-step mechanism where first nucleotides 2–6 of the seed are pre-organized by the MID and PIWI domains in a stacked, helical conformation, with nucleotides 2–4 exposed to the solvent⁴¹. This conformation allows for rapid initial binding, which is weak and metastable and only becomes long-lived if the site presents complementarity to nucleotides 7–8 of the guide strand. Otherwise, AGO disengages the site and either laterally diffuses along the mRNA or dissociates from it completely⁵⁸. Thus, the mRNA-binding behavior of a miRNA is similar to that of an RNA-binding protein (RBP) ^58–60.

AGO–target mRNA crosslinking has enabled the identification of non-canonical (not seed-matching) binding sites^61–63, and a recent large scale microarray-based survey in HeLa cells reported sites with minimal seed-pairing⁶⁴. However, the biological function of non-canonical sites, when considered as a whole, has been disputed as these sites imparted no detectable repression in meta-analyses of miRNA and small RNA transfection datasets ⁵⁴. In addition to the seed, the 3’ half of the miRNA can also contribute to target recognition (particularly nucleotides 13–16, which are termed ‘supplemental region’) in a subset of sites^54,65, and can direct miRNA family members with the same seed to different target sites, as recently shown in C. elegans⁶⁶ and in mouse brains⁶³.

AGO–miRNA binding to the 3’ UTR leads to gene silencing through translation repression and mRNA decay[G] ^2,67 (FIGURE 1B). The latter involves recruitment by AGO of a member of the glycine-tryptophan protein of 182 kDa (GW182) protein family (in humans, trinucleotide repeat-containing gene 6A protein (TNRC6A), TNRC6B and TNRC6C) ^68–71. GW182 interacts with polyadenylate-binding protein (PABPC), thereby promoting mRNA deadenylation by recruiting the poly(A) nuclease 2 (PAN2)–PAN3 and carbon catabolite repressor protein 4 (CCR4)–NOT complexes^72–76. Deadenylation promotes decapping by the mRNA-decapping enzyme subunit 1 (DCP1)–DCP2 complex⁷³, thereby making the mRNA susceptible to rapid degradation by 5’−3’ exoribonuclease 1 (XRN1) ⁷⁷. GW182-mediated recruitment of CCR4–NOT also leads to translation repression through recruitment of the RNA helicase DDX6^2,78.

Inhibition of translation initiation is caused also by interfering with the function of eukaryotic initiation factor 4A-1 (eIF4A1) and eIF4A2^2,79–81. In Drosophila melanogaster S2 cell lysates this activity was independent of GW182⁸⁰, suggesting the existence of multiple mechanisms of translation inhibition. There is disagreement on the mechanism of interference with eIF4A1 and eIF4A2 function^79–81, but most data indicate that miRISC induces their dissociation from target mRNAs and thereby inhibits ribosome scanning and assembly of the eIF4F translation initiation complex. A recent investigation in human and D. melanogaster cells reported the AGO Trp-binding pockets that mediate GW182 binding^41,82 are required for translation inhibition⁸³. Thus, miRISC-mediated translation inhibition is still incompletely understood.

Both modes of miRISC-mediated gene silencing are thought to be interconnected², and ribosome profiling assays revealed that mRNA decay is generally responsible for 66–90% of silencing^67,84. The observation that translation inhibition can be rescued⁸⁵ but that mRNA degradation is irreversible, raises the possibility that the regulated pauses, or blocks, in the molecular cascade leading to mRNA degradation could allow translation repression without mRNA decay.

One miRNA can silence hundreds of genes, although the effect on each gene is generally mild⁸⁶, and multiple miRNAs can regulate the same gene⁸⁷ (FIGURE 1C). Furthermore, entire cellular pathways can be regulated by individual miRNAs⁸⁷ or miRNA clusters[G] ⁸⁸. miRNA binding of neighboring target sites on a target mRNA can result in cooperative repression^65,89,90, which might explain why the function of non-canonical sites can depend on the occupancy of neighboring canonical sites⁹¹. Cooperativity is explained, in part, by the formation of multivalent interactions of GW182 with AGO proteins⁸². miRNAs can either switch off, or function as fine tuners of protein expression⁹², and thereby buffer against noise in gene expression⁹³.

Modification of miRNA sequence

The interaction of miRNAs with their targets is largely based on their seed sequence ^38,54, and miRNA biogenesis if affected by RNA secondary structures that mediate interactions with RBPs⁹⁴. Cellular processes that alter the sequence of a miRNA or of its precursor therefore can affect miRNA biogenesis, activity and turnover.

Isomir formation is a regulated process that determines miRNA activity

Isomirs are mature-miRNA variants that differ in length, sequence or both^31,95–97. The maturation, stability and turnover, activity or targetome of isomirs can vary (FIGURE 2). Isomirs are classified into 5’, 3’ or polymorphic (internal), depending on the site of variation, and can be derived from alternative Drosha or Dicer processing, RNA editing or non-templated nucleotide addition (NTA) ^98,99. A recent study in a human breast cancer cell line reported that many miRNAs had no isoforms whereas miR-21–5p had 43 isoforms¹⁰⁰. Isomir expression patterns have been found useful in distinguishing different types of breast cancer and represent potential biomarkers¹⁰¹.

Figure 2: Isomirs differ in length and sequence and expand the functional repertoire of miRNAs.

Isomirs are classified as 5’, 3’ or polymorphic, with combinations possible. Depending on the arm of the miRNA precursors (5p or 3p; see inset) used to produce the mature miRNA, cleavage by either Drosha or Dicer can result in the formation of the isomir ⁹⁸. 5’ isomirs have shifted seeds and can thereby target a different set of genes¹⁰⁹. The functions of 3’ isomirs are less clear, but there is increased evidence for their differential activity^113,114. Polymorphic isomirs are generated by RNA editing, mainly by ADAR. The editing can affect miRNA biogenesis, either by preventing it, or by leading to the formation of 5’ isomirs or 3’ isomirs; if editing alters the seed, it could retarget a miRNA to other mRNAs^98,99. nt, nucleotides.

The primary determinant of miRNA length and sequence is cleavage by Drosha and Dicer³² (Supplementary Table S1). Alternative processing by Drosha was first reported for miR-142 and miR-342 in mouse CD8 T cells¹⁰² and shown to be widespread in human cells⁹⁷. Cleavage by Dicer has been more extensively studied; it is modulated by TAR RNA-binding protein (TARBP; also known as TRBP) and protein activator of the interferon-induced protein kinase (PACT) in vertebrates (Box 1) and by loquacious in flies^103–106. Variation in the Dicer cleavage site directly modulates the seed sequence of 3p miRNA strands, and can alter guide-strand selection^106,107 (FIGURE 2; Supplementary Table S1). Dicer binds TARBP and PACT by the same domain, as revealed by a recent crystal structure of the binding interface, suggesting Dicer is regulated by a mechanism based on binding–competition between TARBP and PACT, which produce specific isomirs for some miRNAs¹⁰⁶.

Alternative cleavage by Drosha or Dicer can shift the position of the seed and is the primary mechanism of 5’ isomir production (BOX 1; FIGURE 2), as nucleotide addition or removal at the 5’ end of mature miRNAs is rare⁹⁸. Cleavage-directed 5’ end variation can retarget a miRNA, as observed in flies¹⁰⁴ and humans^108–111, and 5’ isomirs can target the same biological pathways together^109,112 (Supplementary Table S1). The abundance of 5’ isomirs can vary between cell types, with certain isomirs being predominant in specific cell types⁹⁶. By contrast, 3’ isomirs mostly differ in their stability and turnover (see below), although recent findings point also to length-related effects on targeting and activity. For example, binding to the RNA of Hepatitis C virus (see below) of a 21-mer 3’ isomir of miR-122 is weaker compared to binding of longer miR-122 isomirs¹¹³, and longer 3’ isomirs of miR-222 have increased apoptotic activity¹¹⁴.

Editing the sequence of miRNA precursors

RNA editing can change miRNA sequence, generating isomirs, and can also affect biogenesis, leading to 5’ or 3’ variation^98,99 (BOX 1; FIGURE 2). Deamination is the most commonly observed type of miRNA precursor editing. Adenosine deaminase acting on RNA (ADAR) converts adenosine to inosine⁹⁹, which is read as guanosine during splicing and translation. Similarly, cytidine deaminase acting on RNA (CDAR) converts cytidine to uracil¹¹⁵. In vertebrates ADAR1 and ADAR2 act on double-stranded RNA and edit miRNA precursors⁹⁹. A-to-I editing sites appear widespread¹¹⁶, and in the human brain ~16% of pri-miRNAs were predicted to be edited (based on the analysis of 209 pri-miRNAs). However, editing levels at different miRNAs vary considerably^99,117.

A-to-I editing of miRNA precursors can interfere with cleavage by Drosha ^117–121 or Dicer ^117,122, or, less frequently, with AGO loading, and editing of the seed can redirect a miRNA to a new set of targets^{120,123–126}. A-to-I editing of miRNAs is involved in miRNA deregulation in cancer^121,127,128. For example, in melanoma, overexpression of cAMP-responsive element-binding protein (CREB) reduces ADAR1 expression, thereby reducing the editing of miR-455–5p; unedited miR-455–5p represses cytoplasmic polyadenylation element-binding protein 1 (CPEB1) and other tumor suppressors, thereby promoting tumor growth and metastasis¹²¹. Changes in seed sequences can also be mediated by the more elusive C-to-U editing¹²⁹. miRNA A-to-I editing was first observed in pre-miR-22 in human and mouse tissues at species-specific positions, and was predicted to alter miRNA targeting and activity¹³⁰. In summary, RNA editing is a versatile regulatory mechanism that can control the abundance and target specificity of miRNAs, but in light of the relatively low frequency of editing⁹⁹, further research is required to characterize its biological importance.

Non-templated nucleotide addition

Non-templated addition of nucleotides (NTA) was first suggested to be biologically regulated during D. melanogaster embryogenesis¹³¹. NTAs predominantly comprise of adenylation or uridylation at 3’ ends, and are miRNA-specific across tissue types, developmental stages, disease states and different species ¹³².

NTAs are carried out by several enzymes and can modulate miRNA stability (FIGURE 3A). Poly(A) RNA polymerase Germ Line Development 2 (GLD2)-mediated 3’-end mono-adenylation stabilizes miR-122 in the liver¹³³ and in mouse embryonic fibroblasts (MEF) ¹³⁴. Poly(A)-specific ribonuclease (PARN) counteracts this as its depletion in a human liver cancer cell line leads to the appearance of miR-122 with 3’-oligo adenylation and to increased miR-122 stability. CUG triplet repeat RNA-binding protein 1 (also known as CELF1) binds miR-122 and mediates its degradation by recruiting PARN (FIGURE 3B). A similar mechanism regulates miR-93 and miR-652–3p¹³⁵. By contrast, a recent analysis of miRNAs in the hyppocampus of GLD2 knockout mice revealed reduced levels of 3’ adenylation, but no effect on miRNA stability¹³⁶; similarly, only about half of the investigated miRNAs were destabilized in human fibroblasts when GLD2-mediated adenylation was suppressed ¹³⁷. Moreover, 3’ adenylation by the poly(A) RNA polymerase PAP-associated domain-containing protein 5 (PAPD5) initiates degradation of miR-21¹³⁸.

Figure 3: Non-templated nucleotide addition and miRNA turnover.

(A) Non-templated nucleotide addition (NTA) by GLD2 can stabilize some miRNAs, but has no effect on others. Terminal 3’ adenylation of miR-21 by PAP-associated domain-containing protein 5 (PAPD5) was even found to promote exonucleolytic cleavage by poly(A)-specific ribonuclease (PARN) ¹³⁸. Terminal 3’ uridylation by terminal uridylyltransferase 4 (TUT4) can reduce the activity of miR-26b¹³⁹, or prime miRNAs for degradation following T-cell activation¹⁴⁰.

(B) Terminal 3’ adenylation stabilizes some miRNAs by counteracting 3’−5’ exonucleolytic activity by PARN. CUG triplet repeat RNA-binding protein 1 (CUGBP1) interacts with miR-122 and recruits PARN¹³⁵.

(C) Interactions of an Argonaute (AGO)-bound miRNA with a target mRNA through the miRNA seed sequence result in translation repression and mRNA degradation. Targets with extensive pairing to the 3’ end promote tailing^144,151 and trimming¹⁴⁸, and target-directed miRNA degradation^142,149,150. A target mRNA that is fully complementary to a miRNA is cleaved when bound by a catalytic Argonaute, such as mammalian Ago2^44,49, but such pairing can also result in unloading of the miRNA from AGO¹⁵⁴. Cell-lines: A549 (adenocarcinomic human alveolar basal epithelial cell line), Huh-7 (human hepatocellular carcinoma cell line), human cellular carcinoma, THP1 (human leukemic monocyte cell line).

Compared to 3’ adenylation, 3’ uridylation more consistently negatively regulates miRNA activity. Uridylation by terminal uridylyltransferase 4 (TUT4; also known as ZCCHC11) of miR-26b in a human adenocarcinoma cell line reduces target repression¹³⁹. Moreover, TUT4-mediated uridylation primes numerous miRNAs for degradation during T-cell activation¹⁴⁰ (FIGURE 3A).

To date, the enzymes reported to mediate NTA include GLD2^133,141, PAPD5¹⁴¹, poly(A) polymerase γ and poly(A) RNA polymerase, mitochondrial¹³² (adenylation) TUT7, TUT4¹³⁹, TUT1 (uridylation). Species-dependent differences in NTAs have been observed, with uridylation being more common in C. elegans and adenylation more common in mouse and human¹³². In summary NTA appears to influence, both positively and negatively, the stability of specific miRNAs in specific cell types. Recent findings demonstrate that interactions between the miRNA 3’ region and target RNAs can substantially affect miRNA modification and turnover (see below), thereby providing a potential mechanism for the cellular-context-dependence of NTA.

miRNA turnover

miRNAs are generally thought to be stable in vivo, but turnover rates are regulated by multiple factors and across different miRNAs can range from minutes to days¹⁴². Turnover rates are miRNA-specific and isomir-specific and thus linked to miRNA sequence^143,144; indeed, a 5’ guanine or cytosine is associated with faster turnover rates compared to a uracil in this position¹⁴³. Stability can also be miRNA-specific. For example, miRNAs in mouse fibroblast 3T3 cells are generally stable, except for miR-16 family members, which are intrinsically unstable¹⁴⁵. This instability allows miR-16 levels to vary with, and thereby help regulate, the progression of the cell cycle. Some miRNAs are stable unless specifically degraded in response to developmental cues; for example, miR-150 is rapidly lost when murine naive T cells differentiate into Th1 and Th2 lymphocytes¹⁴⁶. Turnover rates can depend also on tissue context: fast turnover rates are generally observed in neuronal miRNAs compared to miRNAs in other tissues¹⁴⁷.

One mechanism of miRNA destabilization is termed target RNA–directed miRNA degradation (TDMD) ¹⁴⁸. In TDMD, target RNAs with extensive complementarity to both 5’ and 3’ regions of a miRNA promote its turnover (FIGURE 3C). TDMD is often associated with 3’ NTA, or “tailing”, and with 3’−5’ trimming of the degraded miRNA^142,149,150. It has been proposed that 3’ tailing extends the AGO-loaded miRNA^144,151 until a 3’−5’ nuclease can bind and degrade it¹⁴⁸. This hypothesis is consistent with the observation that 3’-remodeled isomirs generally coincide with TDMD. However, the hypothesis has recently been challenged by the observation that tailing of miR-7 by GLD2, which coincides with TDMD of miR-7 in mouse neurons, is not required for efficient miR-7 degradation¹⁵². Similarly, the 3’−5’ exoribonuclease DIS3-like exonuclease 2 (DIS3L2) was implicated in 3’ trimming associated with TDMD of miR-27a, but DIS3L2 depletion did not impact TDMD efficiency¹⁵¹. Thus, it is possible that TDMD and miRNA 3’ remodeling are distinct processes, and the molecular mechanism of TDMD therefore remains poorly understood.

Although viral inducers of TDMD have been known for several years (see below), endogenous inducers of TDMD have only recently been reported. In zebrafish, miR-29b is selectively degraded in the cerebellar granule cell layer by a long non-coding RNA (lncRNA)and in mice by the 3’ UTR of a protein-coding gene¹⁵³. Both transcripts show extensive sequence similarity around a highly complementary target site for miR-29b. Similarly, the lncRNA Cyrano can induce TDMD of miR-7, thereby affecting the stability of a miR-7-interacting circular RNA (circRNA) (see below) ¹⁵². Furthermore, miR-503, which is inherently unstable in 3T3 cells, can be stabilized by mutating either its seed or 3’ regions¹⁴⁵, raising the possibility that its inherent instability is mediated by TDMD. It is possible that many inherently unstable miRNAs are subjected to TDMD through yet-to-be-identified target RNAs.

A biochemical screen using HEK239 cell (a human embryonic kidney cell line) lysate revealed that miRNA degradation is promoted particularly by seedless targets with high complementarity to miRNA 3’ ends, with complementarity of the 3’-terminal nucleotide being crucial for degradation¹⁵⁰. Combined with the observation that the sequences of both the 5’-end and the 3’-end influence turnover rates^143,144, this might partially explain the tissue-specific and miRNA-specific effects of NTA on stability. Alternatively, highly complementary targets can destabilize the AGO–miRNA interaction and promote release of the guide strand¹⁵⁴, although miRNAs appear to be stabilized by high abundance of seed-matching targets ¹⁵⁵.

In addition to exonucleolytic miRNA turnover, Tudor staphylococcal nuclease, which is responsible for degradation of A-to-I edited double-strand RNA⁹⁹, was found to target miR-31–5p, miR-29b-3p and miR-125a-5p in HEK293T cells. Other tested miRNAs did not appear to be affected, suggesting this is a miRNA-specific mechanism, though how selectivity is established remains to be determined¹⁵⁶. Finally, phosphorylation-dependent regulation of miRNAs was recently discovered for the tumor suppressor miR-34. The nuclei of four human cancer cell lines were found to contain pools of inactive, mature single-stranded miR-34 lacking a 5’ phosphate and not bound by AGO¹⁵⁷. Irradiation of the cells led to 5’ phosphorylation, export to the cytoplasm and loading into AGO of miR-34. The serine-protein kinase ATM and the RNA kinase Clp1 were required for miR-34 phosphorylation. It is unclear how widespread this phenomenon is, but these findings suggest that miRNAs can be maintained in inactive form and rapidly activated in response to stimuli.

Post-translational modification of AGO

AGO proteins undergo PTM on multiple residues (FIGURE 4). Phosphorylation is the best characterized AGO PTM and occurs at three main sites: at the L2 linker, on Ser387 and Tyr393 of human Ago2; at the miRNA 5’-end binding region in the MID domain, on Tyr529; and at a surface-exposed loop in the PIWI domain known as the eukaryotic insertion or the S824–S834 cluster, which includes Ser824, Ser828, Thr830, Ser831 and Ser834.

Figure 4: The activity and the stability of miRISC is modulated by post-translational modifications (PTMs) of Argonaute proteins.

(A) Phosphorylation of Ser387 (S387) in the L2 region of Argonaute (AGO) was found to be mediated by MAP kinase-activated protein kinase 2 (MAPKAPK2) ¹⁵⁸ and AKT3¹⁵⁹ in vitro. Ser387 phosphorylation increases miRNA activity by stimulating the assembly of miRNA-induced silencing complexes, and reduces translocation of Ago2 to multivesicular endosomes and secretion of exosomes¹⁶². Phosphorylation of the nearby Tyr393 (Y393), also in the L2 region, decreases the miRNA–Ago2 association, thereby reducing miRNA activity^164,165. Tyr 529 (Y529) is located in the middle (MID) domain, near the miRNA 5’-nucleotide binding pocket, and its phosphorylation prevents miRNA loading¹⁶⁰. No function has yet been assigned to phosphorylation sites in the Piwi-Argonaute-Zwille (PAZ) domain (S253, T303, T307) and the P-element induced wimpy testes (PIWI) domain (S798) ¹⁶⁰. Additional AGO PTMs include Pro700 (P700) 4-hydroxylation, which increases Ago2 stability¹⁷¹. Lys402 (K402) SUMOylation, which was reported to either destabilize Ago2¹⁷⁷ or be required for full siRNA activity¹⁷⁸; and poly(ADP-ribosylation) (PARylation), which inhibits miRNA activity^173,174, presumably by decreasing target accessibility.

(B) The S824–S834 cluster in the eukaryotic insertion region of human Ago2 undergoes a phosphorylation cycle, which regulates AGO–target interactions. Phosphorylation of the Ser residues in the cluster by casein kinase I isoform α (CSNK1A1) favors target release. Subsequent dephosphorylation by the serine/threonine-protein phosphatase 6 complex ANKRD52–PPP6C primes Ago2 for the next round of target binding^169,170.

Several studies report phosphorylation of Ser387 in the L2 region of human Ago2 (FIGURE 4A) ^158–160. Ser387 phosphorylation is stimulated by the p38 MAPK pathway-mediated stress response, for example by MAP kinase-activated protein kinase 2 in vitro¹⁵⁸. The proto-oncogene serine/threonine kinase AKT3 also phosphorylates Ser387 in vitro and in HeLa cells and its depletion or the expression of a S387A Ago2 mutant (which cannot be phosphorylated) led to de-repression of a luciferase reporter and weakened the Ago2 interactions with TNRC6A¹⁵⁹. Similarly, Ser387 phosphorylation is necessary for interactions between Ago2 and LIM domain-containing protein 1, which is a protein suggested to bridge Ago2 and GW182¹⁶¹. Blocking Ser387 phosphorylation also reduced accumulation of Ago2 in cytoplasmic foci with GW182^158,159 and was suggested to reduce trafficking of Ago2–miRNA complexes to multivesicular endosomes[G] and reduce the secretion of exosomes[G] ¹⁶². Notably however, a recent investigation found that S387A Ago2 did not reduce co-localization GW182 in HeLa cells, indicating that additional factors may be involved in enabling the Ago2–GW182 interaction¹⁶³. The combined data indicate that Ser387 phosphorylation promotes miRNA function by stimulating the assembly of miRISC.

Tyr393 is adjacent to Ser387 in the L2 region and its phosphorylation is also well-documented^160,164,165 (FIGURE 4A). It appears to be mediated by epidermal growth factor receptor (EGFR) and stimulated by hypoxic stress¹⁶⁴. Overexpression of oncogenic RAS reversibly inhibits protein tyrosine phosphatase 1B (also known as tyrosine-protein phosphatase non-receptor type 1), resulting in Ago2 hyperphosphorylation at Tyr393¹⁶⁵. Both studies reported diminished interactions between Ago2 and Dicer and reduced levels of Ago2-associated miRNAs following Tyr393 phosphorylation. Thus, in contrast to Ser387, Tyr393 phosphorylation appears to negatively regulate miRNA activity by inhibiting the loading of miRNAs into Ago2, thereby promoting tumorigenesis.

Tyr529 phosphorylation also potentially blocks miRNA loading¹⁶⁰ (FIGURE 4A). Tyr529 binds the 5’ phosphate and first nucleotide of miRNAs loaded into AGO^41,166,167 Tyr529 phosphorylation is therefore expected to preclude miRNA binding, as indeed was demonstrated in Tyr529 Ago2 mutants¹⁶⁰. Tyr529 phosphorylation was also associated with decreased miRNA binding to Ago2 during macrophage activation¹⁶⁸.

A phosphorylation cycle of residues S824–S834 in Ago2 regulates the release of target mRNAs^169,170 (FIGURE 4B). Target-binding leads to phosphorylation of these residues by casein kinase I isoform α, which reduces the affinity of Ago2 for mRNA and enables target release. The serine/threonine-protein phosphatase 6 complex ANKRD52–PPP6C dephosphorylates the residues, which primes Ago2 for binding a new target. Interrupting this cycle strongly inhibited miRNA activity, and the cycle was suggested to prevent overly long association with mRNA targets¹⁷⁰.Alternatively, the phosphorylation cycle may represent an AGO-regulation mechanism that is mediated by mRNA-binding proteins. The cluster is conserved in human AGO proteins, mouse Ago2, rat Ago2, zebrafish Ago2, the C. elegans argonaute ALG1¹⁶⁹, and in the fruit fly Ago1 (but not in the fruit fly Ago2, which is used for siRNA-mediated target cleavage) ¹⁷⁰.

Additional Ago2 phosphorylation sites have been reported in the PAZ domain (Ser253, Thr303, Thr307) and the PIWI domain (Ser798), but their function is currently unknown¹⁶⁰ (FIGURE 4A).

In addition to phosphorylation, other AGO PTMs are known (FIGURE 4A). Hydroxylation of Ago2 Pro700 appears to increase the stability of Ago2 in mouse and human cells¹⁷¹. Interestingly, Pro700 is adjacent to a Trp-binding pocket in Ago2 that is used to interact with GW182^41,82, and its 4-hydroxylation thus could potentially alter the assembly of miRISC, though this has not yet been demonstrated.

Ago1–4 undergo poly(ADP-ribosylation) (PARylation) in human cell lines¹⁷². Poly(ADP-ribose) (PAR) is a polymer of ADP-ribose units that can be enzymatically added to Asp, Glu and Lys residues. PARylation regulates the cellular stress response, particularly the formation of cytoplasmic stress granules, which regulate mRNA translation and stability¹⁷². PARylation of Ago1–4 reduced translation repression and endonucleolytic cleavage and was hypothesized to impair target accessibility¹⁷³. PARylation-induced down regulation of miRNA activity was observed in antiviral response in HEK293 cells: cytotoxic interferon-stimulated genes are generally heavily regulated by miRNAs, and miRNA inhibition through AGO PARylation might enable a cytotoxic response to viral infection¹⁷⁴.

The ubiquitin–proteasome system reduces the abundance of the fruit fly Ago1 and mouse Ago2¹⁷⁵, and drives global miRNA down-regulation during T cell activation¹⁷⁶. Finally, two groups have reported Ago2 SUMOylation at Lys402, with contradicting consequences: SUMOylation might destabilize Ago2¹⁷⁷, or be required for full Ago2 siRNA-mediated activity¹⁷⁸.

Compared to AGO proteins, less is known about the regulation of other proteins involved in miRNA function. Tripartite motif-containing protein 65 ubiquitylates TNRC6 in HEK293 cells, leading to proteasomal degradation and de-repression of miRNA targets¹⁷⁹. TNRC6A has long been known to be phosphorylated in mammalian cells¹⁸⁰, and phosphorylation of the PAM2 motif [G] of TNRC6C was suggested to reduce its interactions with PABPC1¹⁸¹. A recent mass-spectrometry study of the TNRC6A interactome in HeLa cells revealed a large number of binding partners, but it is currently unknown whether any of these regulate TNRC6A¹⁸².

Regulation of miRNAs by sequestration

The competing endogenous RNA (ceRNA) hypothesis ¹⁸³ postulates that an increase in the cellular concentration of a miRNA target RNA would reduce the cytoplasmic availability of the specific miRNA by binding it, thereby de-repressing other mRNAs that are targets of the same miRNA; this is similar to the proposed function of miRNA sponges[G] ¹⁸⁴. Thus, according to the ceRNA model, gene expression would be shaped by the global competition of target RNAs for miRNAs¹⁸⁵ (FIGURE 5A). The potential efficacy of ceRNAs is controversial, as the increased expression of any single ceRNA is expected to increase only slightly the total number of target sites of a miRNA, and thus be unlikely to meaningfully affect miRNA activity^186–189. Two studies measuring the cellular abundance of miRNAs and miRNA target sites proposed that the ability of a putative ceRNA to de-repress the expression of other target mRNAs depends on miRNA abundance¹⁹⁰ and/or target site abundance¹⁸⁶. Both studies agreed that, in most cases, unphysiologically high levels of a ceRNA would generally be necessary to yield a biologically meaningful effect^186,190. Mathematical modeling of miRNA distribution in the targetome indicates that a ceRNA must increase the cellular abundance of target sites at least two fold to be effective¹⁸⁸. On the other hand, it was also noted that ceRNA efficacy could be enhanced if local concentrations of ceRNAs and/or miRNAs deviated strongly from the cytoplasmic average¹⁸⁸. The discovery of the phosphorylation cycle regulating AGO activity also raises interesting new possibilities, as AGO–target binding might be modulated by mRNA-associated RBPs¹⁷⁰. With respect to the ceRNA hypothesis, such a mechanism could alter the affinities of different target sites that otherwise seem equivalent based on miRNA-complementarity alone. Additionally, although this has yet to be demonstrated, ceRNAs that promote TDMD could function at lower concentrations by reducing miRNA abundance¹⁸⁹. Thus, although the original ceRNA hypothesis may be implausible under general cellular conditions, ceRNA mechanisms may still have a biological role ¹⁸⁸.

Figure 5: miRNA sequestration by endogenous and viral RNAs.

(A) The competing endogenous RNA (ceRNA) hypothesis states that a newly expressed RNA can compete with the already present microRNA (miRNA) targets for cytoplasmic miRNA-induced silencing complexes (miRISCs), potentially leading to de-repression of certain genes¹⁸⁵. However, a ceRNA is unlikely to lead to gene de-repression when it is expressed at typical physiological levels^186–190.

(B) Long non-coding RNAs (lncRNAs), pseudogenes and mRNAs can have ceRNA activity. In mice, long intergenic non-protein coding RNA of muscle differentiation 1 (Linc-md1) contains one binding site for miR-133 (blue) and two for miR-135 (green). By sequestering these miRNAs, the muscle-specific transcription factors mastermind-like protein 1(Maml1) and myocyte-specific enhancer factor 2C (Mef2c) are de-repressed, thereby promoting myoblast differentiation¹⁹¹. The pseudogene phosphatase and tensin homolog pseudogene 1 (PTENP1) shares many miRNA target sites with the tumor suppressor PTEN, and can de-repress PTEN in human cells ¹⁸³. Similarly, the mouse zinc finger E-box binding homeobox 2 (Zeb2) mRNA can de-repress PTEN¹⁹³.

(C) Different viral mechanisms affect miRNA function. Hepatitis C virus (HCV) harbors two binding sites for miR-122 at the very end of the 5’ untranslated region (UTR) of its RNA genome²⁰³. These recruit argonaute2 (AGO2)–miR-122 complexes²⁰⁴ to protect the viral RNA from the cellular antiviral response and the activity of exonucleases²⁰⁵, and functionally sequester miR-122 and de-repress hepatic miR-122 target mRNAs²⁰⁷.

Bovine viral diarrhea virus (BVDV) is an RNA virus that contains a binding site for miR-17 in its 3’ UTR; miR-17 binding enhances the stability of the viral RNA. The site functionally sequesters the miRNA and de-represses its cellular targets²⁰⁸.

Herpesvirus saimiri (HVS) produces short non-coding RNAs termed herpesvirus saimiri uracyl-rich RNAs (HSURs), two of which are known to modulate miRNA function. HSUR1 binds miR-27a and promotes its TDMD, which leads to de-repression of miR-27a cellular targets and promotes T cell activation²¹². HSUR2 binds miR-142–3p and miR-16 and tethers them to cellular target mRNAs, which prevents apoptosis²¹⁴.

The controversy notwithstanding, the list of potential ceRNAs is growing and includes lncRNAs, pseudogenes, mRNAs (FIGURE 5B) and specific circRNAs¹⁸⁷. In mice, long intergenic non-protein coding RNA of muscle differentiation 1 (Linc-md1) is proposed to drives myoblast differentiation by sequestering miR-133 and miR-135, which target muscle-specific transcription factors¹⁹¹. Similarly, in human cells the untranslated pseudogene phosphatase and tensin homolog pseudogene 1 (PTENP1) de-represses PTEN¹⁸³. This was suggested to be a common function of pseudogenes¹⁸⁷, as they regularly share miRNA target sites with their parent genes. Recently the overexpression of the Braf pseudogene was found to promote B cell lymphoma in mice through a possible ceRNA mechanism¹⁹². mRNAs have also been suggested to function as ceRNAs; for example, PTEN expression is regulated not just by pseudogenes, but also by numerous protein-coding ceRNAs^193,194 (FIGURE 5B). Different types of ceRNAs were found to crosstalk in a network of various oncogenic pathways in glioblastoma¹⁹⁵.

circRNAs are enigmatic ncRNAs with dynamic and complex expression patterns, whose function is still poorly understood¹⁹⁶. The circRNA CDR1 antisense RNA (CDR1as; also known as ciRS-7), is expressed in human and mouse brains^197,198. Because CDR1as contains many binding sites (63–74) for miR-7, it acts as a miR-7 sponge when overexpressed¹⁹⁷. In support of the sponge mechanism, depletion of CDR1as in HEK293 cells led to downregulation of miR-7 target mRNAs¹⁹⁸. However, knockout of CDR1as in mice led to reduction in miR-7 levels and increase in the levels of miR-7 target genes in the brain, suggesting the sponge effect stabilizes miR-7, as none of its binding sites in CDR1as possesses the extensive 3’ complementarity required for TDMD¹⁹⁹. The miR-7 molecules can be released from CDR1as through its slicing by miR-671, for which CDR1as contains a highly-complementary target site. CRD1as may thus bind to miR-7 in order help to localize it to specific subcellular compartments^197–199. The lncRNA Cyrano was proposed to be responsible for the increased turnover of miR-7 upon CDR1as depletion¹⁹⁹, and has now been reported to induce TDMD of miR-7¹⁵². Interestingly, Cyrano depletion increased miR-7 levels and led to a decrease in CDR1as levels, in part due to an enhanced miR-671-mediated degradation of CDR1as¹⁵².

Although thousands of circRNAs have been identified, only a handful stand out as harboring large numbers of miRNA target sites²⁰⁰. In addition to CDR1as, ten different circRNAs derived from zinc finger genes contain 7–24 target sites for miR-23, miR-181 or miR-199 families²⁰⁰. Another proposed sponge, the testis-specific circRNA sex-determining region Y, possesses 16 sites for miR-138 in mice¹⁹⁷, but only one in humans²⁰⁰. Thus, whereas some circRNAs may interact with miRNAs, this does not appear to be a major function of most of them.

Viral modulation of miRNA activity

Viruses have evolved to repurpose or modulate host miRNAs for their replication, which often affects miRNA function (FIGURE 5C). The best studied interaction is between the Hepatitis C virus (HCV) and miR-122. HCV is a positive-sense single-stranded RNA virus of the Flaviviridae family, which causes acute and chronic liver infection²⁰¹. Its highly structured 5’ UTR contains an internal ribosomal entry site²⁰², and upstream of it, at the very 5’ end of the genomic RNA, two binding sites for the liver-abundant miR-122²⁰³ (FIGURE 5C). The binding sites recruit Ago2–miR-122 to the uncapped 5’ end of the viral RNA²⁰⁴ to protect it from the cellular antiviral response and exonuclease activity^205,206. This functionally sequesters cytoplasmic miR-122, which is a tumor suppressor miRNA responsible for hepatic maintenance, and potentially explains the link between chronic HCV infection and an increased risk of developing hepatocellular carcinoma²⁰⁷.

Another virus from the Flaviviridae family that binds host miRNAs is bovine viral diarrhea virus, which is an economically important cattle virus. Binding of let-7 and miR-17 at a structured region of the viral RNA 3’ UTR promotes viral replication. miR-17 also stabilizes the viral RNA upon binding and enhances its translation through an unknown mechanism. The binding of miR-17 to the viral RNA de-represses cellular targets of the miRNA, an effect also observed with classical swine fever virus²⁰⁸. The detection of similar interactions between miRNAs and viral genomic RNA of different members of the Flaviviridae family raises the question of whether this mechanism might be even more widespread in this family, which includes other health-relevant members such as dengue virus and zika virus.

Hepatitis B virus (HBV), which is a partially double-stranded DNA virus unrelated to HCV, directly down-regulates miR-122 in liver cells, as this miRNA appears to inhibit viral infection²⁰⁹. A hybrid viral–human transcript²¹⁰ contains multiple miR-122 binding sites, one of which resembles a TDMD element that depletes cellular miR-122 levels and promotes loss of hepatic function and liver damage in mice²¹¹.

Several members of the Herpesviridae family also modulate host miRNA function. Herpesvirus saimiri (HVS) expresses H. saimiri uracyl-rich RNAs (HSURs), which are short ncRNAs with structural similarity to small nuclear RNAs[G]. Two HSURs harbor binding sites for host miRNAs: miR-16 (HSUR2), miR-27a (HSUR1) and miR-142–3p (both HSUR1 and HSUR2). HSUR1 reduces miR-27a levels in infected marmoset T cells through TDMD, thereby de-repressing miR-27a cellular target mRNAs and promoting T cell activation²¹². TDMD of miR-27a is required for efficient HVS replication, as viral strains with HUSR1 bearing a mutated miR-27a binding site have reduced titers²¹³. HSUR2 does not deplete the miRNAs it binds, but instead acts as a tether that recruits AGO–miR-142–3p and AGO–miR-16 complexes to cellular mRNAs that encode pro-apoptosis factors, thereby inducing silencing of the tethered mRNAs and preventing apoptosis²¹⁴.

Analogously to HSUR1 function, TDMD and reduction in miR-27a levels was observed in mouse cell lines and primary macrophages upon murine cytomegalovirus infection²¹⁵, mediated by a transcript harboring a miR-27a target site with substantial 3’-end complementarity²¹⁶. Similarly, human cytomegalovirus (HCMV) targets miR-17 and miR-20a, two members of the miR-17–92 cluster. Degradation of these miRNAs, which is mediated by complementarity with sites in viral non-coding RNA, accelerates virus production during HCMV infection²¹⁷.

miRNA transport from the cytoplasm

As miRNAs function in the cytoplasm, their activity can be modulated by transferring them to the nucleus or to extracellular vesicles, where they potentially have location-specific functions²¹⁸.

Nuclear localization of miRNAs

miR-21 was the first miRNA to be found in the nucleus, as 20% of total miR-21 is present in nuclear extracts⁴⁹; miR-29b was found to be predominantly localized to the nucleus owing to a 3’ hexanucleotide nuclear localization signal²¹⁹. Ago2, Dicer, TARBP and GW182 are present in the nucleus and form complexes. In C. elegans, the AGO homolog ALG-1 uses mature let-7 to bind pri-let-7 transcripts in the nucleus and promote their biogenesis, thereby creating a positive feedback loop²²⁰. Recent single molecule studies in mammalian cells indicated that nuclear miRNAs do not repress complementary targets. The importance of the miR-29b hexanucleotide localization signal was also questioned, as miRNA nuclear localization was found to be primarily dependent on the presence of nuclear targets¹⁵⁵. These findings suggest that our understanding of nuclear miRNA activity is still very partial.

Circulating miRNAs

Circulating miRNAs[G] are potential cancer biomarkers²²¹; their discovery in exosomes led to the hypothesis that they might contribute to inter-cellular signaling²²² (FIGURE 6A). It remains unclear if only a small fraction of circulating miRNAs travel within exosomes (~10% or less in plasma) ^223,224, or if exosomes contain the majority (83–99% in serum) ²²⁵ of circulating miRNAs. These apparently contrasting observations could be due to technical differences in exosome isolation, or differences between serum and plasma²²⁵.

Figure 6: Mechanisms of sorting miRNAs into exosomes.

(A) miRNAs can be packaged into exosomes and thus may contribute to inter-cellular signaling. Uptake of exosomes can be receptor-mediated or receptor-indepedent²⁷⁸. Upon entering a target cell, exosome-delivered miRNAs are speculated to regulate target mRNAs²²².

(B) Although the biological function of exosomal miRNAs is still incompletely understood, multiple mechanisms direct miRNAs into exosomes. Argonaute 2 (Ago2)-dependent sorting of specific miRNAs has been reported in isogenic colon cancer cells, and phosphorylation of Ago2 Ser387 inhibited loading of some miRNAs into exosomes¹⁶². Sorting based on miRNA-sequence complementary of exosomal long non-coding RNAs (lncRNA) was shown for miR-149–3p in prostate cancer cells²³². Exosomal RNA-binding proteins (RBP) can direct miRNAs into exosomes by binding ‘exomotifs’ at the miRNA 3’ ends. The exomotif GGAG promotes exosomal sorting of miR-198 by heterogeneous nuclear ribonucleoproteins A2/B1 (hnRNPA2B1) in human primary T cells²³³, and the exomotif GGCU promotes hnRNPQ -mediated exosomal sorting of miRNAs in murine hepatocytes²³⁴. No exomotifs are known for Y-box-binding protein 1 (YBX1)-mediated sorting of miR-223 in HEK293T cells²³⁵ and for major vault protein (MVP)-mediated sorting of miR-193a²³⁶. Finally, in human B cells 3’-adenylated miRNAs are depleted in exosomes whereas 3’-uridylated miRNAs enriched in exosomes²³⁷. Exosomes have been reported to carry proteins, different RNAs and miRNA biogenesis components ²²⁶, but also Ago2–miRNA complexes¹⁶² and AGO-free miRNAs²³⁵.

Circulating miRNAs may have regulatory functions. Breast-cancer cell derived exosomes contain pre-miRNAs and the proteins required for cell-independent miRNA maturation, can induce cell proliferation in culture²²⁶, and might have prometastatic properties in vivo^227,228. A major concern with these studies is the technical challenge of clearly distinguishing between miRNA-mediated effects and changes caused by other exosomal components²²⁹. Moreover, exosomes from five different sources were found to have less than one miRNA molecule per exosome on average²³⁰. Nevertheless, a recent study showed that brown adipose tissue is a major source of exosomal miRNAs in humans and mice, in support of a functional role of exosomal miRNAs²³¹. Specifically, exosomes derived from a donor mouse expressing an exogenous miRNA in brown adipose tissue were found to repress a reporter gene in the liver of an acceptor mouse²³¹.

A biological function of exosomal miRNAs would require miRNA-specific mechanisms of sorting into exosomes (FIGURE 6B). One proposed sorting mechanism is through exosomal lncRNAs, based on the observed enrichment in exosomes of prostate cancer cell lines of miR-149–3p with lncRNAs harboring its target sites²³². Exosomal RBPs could also promote sorting by directly interacting with miRNA motifs termed ‘exomotifs’. The exomotif GGAG located at the miRNA 3’ end was suggested to mediate binding of miR-198 and other exosomal miRNAs to the exosomal RBP heterogeneous nuclear ribonucleoproteins A2/B1 (hnRNPA2B1) in human primary T cells, thereby directing the miRNAs to exosomes²³³. Similarly, the exomotif GGCU, which is also preferentially located at miRNA 3’ ends, was suggested to recruit hnRNPQ (also known as SYNCRIP) in murine hepatocytes, yielding exosomal secretion of the bound miRNAs²³⁴. Other RBPs proposed to mediate miRNA sorting into exosomes, such as Y-box-binding protein 1 (also known as nuclease-sensitive element-binding protein 1) for miR-223 in HEK293T cells²³⁵ and major vault protein for miR-193a in colon tumor cells in mice²³⁶, lack specific associated recognition motifs. The tumor suppressor miR-193a is an interesting case, because the main reason for its sorting into exosomes appears to be its removal from the cytoplasm²³⁶. NTA at miRNA 3’ ends may also influence sorting, as reported for human B cells and B-cell-derived exosomes, where adenylation was associated with miRNA depletion from and uridylation with miRNA enrichment in exosomes²³⁷. Finally, Ago2 Ser387 phosphorylation by the KRAS–MAPK pathway antagonizes exosomal sorting of Ago2 in colon cancer cells¹⁶². Although free, single-stranded miRNAs are rapidly degraded in the cell and thus are hardly functional¹⁵⁵, some reported sorting mechanisms^233,234 focus on miRNAs, with no mentioning of AGO proteins, or do not detect AGO proteins in exosomes²³². Clearly, more studies are required to fully understand the mechanistic and functional aspects of exosomal miRNAs.

Regulation of miRNA target sites

The activity of miRNAs can also be modulated through changes in their target sites. RNA editing can alter target site complementarity and deregulation of target site editing is associated with cancer^238,239. Similarly, the t1A binding site on AGO does not recognize N6-methyladenosine modification in RNA, raising the possibility that adenosine methylation of target mRNAs could subtly modulate their targeting by miRNAs, though this idea has yet to be tested in a cellular context⁵⁷. Less subtly, formation of mRNA 3’ UTR isoforms can add or remove miRNA target sites, thereby altering mRNA susceptibility to miRNAs in a cell-type-specific or tissue-specific manner²⁴⁰.

RBPs also can modulate miRNA–target interactions. Pumilio promotes cell-cycle re-entry of quiescent cells by binding the 3’ UTR of the mRNA encoding the tumor suppressor p27 (also known as CDKN1B) and remodeling it, thereby exposing target sites for miR-221 and miR-222²⁴¹. AU-binding factor 1 (AUF1) facilitates interactions of Ago2 with mRNAs in HeLa cells²⁴². The AU-rich element binding protein human antigen R (HuR; also known as ELAV-like protein 1) de-represses the miR-122-targeted mRNA encoding cationic amino acid transporter 1 (CAT1) by binding an AU and U rich region in the mRNA, thereby preventing miRNA function but apparently not miRNA binding⁸⁵. Regulation by HuR appears to be widespread, as over 75% of human mRNAs that harbor miRNA binding sites also possess HuR binding sites, often in overlap or in close proximity²⁴³. HuR prevents repression of p53 by miR-125b by binding to the mRNA and causing miRISC dissociation²⁴⁴, and of programmed cell death protein 4 (PDCD4) by miR-21 by binding to targets in the mRNA and displacing miRISC and by directly binding to miR-21 in MCF cells²⁴⁵. RBPs have also been suggested to regulate the Ago2 phosphorylation cycle, but no specific RBP has been discovered yet to do so¹⁷⁰.

Interestingly, in extracts of most adult mouse tissues, miRNAs are found in low molecular weight miRISC of ~100 kDa, which is approximately the molecular mass of AGO with its miRNA guide and not binding mRNA. By contrast, in cell lines most miRNAs are found in high molecular weight miRISC (up to 2 MDa) containing GW182, other proteins and target mRNAs. Consequently, in transformed cells miRNA abundance may correlate more strongly with miRNA activity than in primary tissues, where additional regulation of AGO activity by the cell appears present²⁴⁶. Findings obtained in cell lines should therefore be considered with caution.

Future perspective

A number of interesting questions remain unanswered regarding the regulation of miRNA function. Non-templated nucleotide addition at the 3’ end of miRNAs has been recognized as a widespread process^131,132. However, neither the tissue-specific significance of the added nucleotides that promote miRNA stability or degradation, nor the mechanism by which the cell selects which miRNAs to modify are known. Moreover, an extended miRNA 3’ end has been shown to improve interactions between miR-122 and HCV genomic RNA¹¹³, and to alter the function of miR-222 to promote apoptosis in a breast cancer cell line¹¹⁴, suggesting that 3’-isomir variation may affect miRNA targeting in ways that remain to be understood. Thus, whereas altered seed-targeting has been explored in 5’ isomirs, very little is known about the functions and properties of 3’ isomirs.

The role of mRNA secondary and tertiary structure in miRNA–target recognition is also largely unexplored. Structures that render target sites inaccessible to miRISC clearly inhibit silencing^247,248. However, miRNA binding sites have been identified within heavily structured segments of RNA viruses^203,208, raising the possibility that some structures may not be detrimental for targeting or might even enable recognition by miRNAs. Indeed, effective miR-159 target sites in Arabidopsis thaliana require an adjoining structural element composed of two stem loops²⁴⁹. The interactions between RNA structures and AGO and the degree to which they shape miRNA targeting remain unknown.

Although the activity of exosomal miRNAs and the presence of miRNA-selective exosomal sorting mechanisms suggest exosomal miRNAs participate in intercellular communication, evidence in a physiological context remains elusive. The reported scarcity of miRNAs in exosomes²³⁰ and the difficulty of disentangling miRNA-mediated effects from effects of other exosomal cargoes add to this challenge²²⁹. Similarly, although exosomal Ago2-loaded miRNAs have been detected^162,226, so were single-stranded miRNAs, which would require loading into AGO in the target cells²³⁵. Our understanding of exosomal miRNAs is thus far from complete.

Finally, the recent report of molecular condensation properties of Argonaute and TNRC6²⁵⁰ connects miRNA regulation to the growing field of biological phase separation²⁵¹. The data demonstrate that miRISC can form large molecular condensates in vitro and in living cells and it was hypothesized that the ability to form higher order complexes through molecular condensation may allow miRISC to organize miRNA–target interactions within the cytoplasm and thereby modulate rates of mRNA translation and decay. This raises the possibility that miRNA activity is regulated through the assembly of miRISC itself, by modulation of the biophysical properties of miRISC components.

In conclusion, although the first evidence of miRNAs was discovered over 25 years ago and major advances have been made since, many aspects surrounding the complex mechanisms that govern the activity of these tiny regulators remain to be discovered and explored.

Glossary

Isomirs

Variant forms of a canonical miRNA, generated by alternative cleavage during biogenesis, RNA editing, or non-templated nucleotide addition

guide strand

The strand in the mature miRNA duplex generated by Dicer cleavage, which is loaded into AGO and used to identify complementary sites in target mRNAs

Metastable

A stable state in a system that is not the state of least energy

mRNA decay

Controlled nucleolytic degradation of mRNA, usually starting with de-adenylation, and then progressing either with 3’−5’ exonucleolytic processing, or de-capping and 5’−3’ exonucleolytic processing

multivalent interactions

Interactions between binding partners mediated by multiple individual, often times relatively weak, binding events or points of contact

multivesicular body

A type of late endosome that contains intraluminal vesicles formed by budding into the lumen of the endosome. Its content can be degraded by fusion with lysosomes, or released into the extra-cellular space through fusion with the cell membrane

exosomes

A type of extracellular vesicles, ~50–150 nm in diameter, which contain proteins, lipids and RNA and can carry cargo to target cells

PAM2 motif

Poly A binding protein interacting motif 2, which mediates the interaction between GW182 and PABP

small nuclear RNAs

Small non-coding RNAs localized to the nucleus that form complexes with proteins and are part of the splicing machinery

circulating miRNAs

miRNAs present in circulation and found either as AGO:miRNA complexes or as vesicle (exosome) cargo

miRNA cluster

Multiple miRNAs located in close proximity on the genome and transcribed as a single pri-miRNA

seedless target

A miRNA target with severely reduced complementarity to the miRNA seed

Trp-binding pocket

Ago possesses three pockets located in the PIWI domain, that bind Trp and mediate the interaction with GW182

Stress granules

Granules formed by untranslating mRNAs, initiation factors, RBPs and other proteins, when translation stops as part of the cellular stress response.

miRNA sponge

A transcript expressed from a strong promoter in a cell, which contains multiple target sites for a specific miRNA, thereby derepressing the miRNAs targets