Post-transcriptional regulation of polycistronic microRNAs

An important proportion of microRNA (miRNA) genes tend to lie close to each other within animal genomes. Such genomic organization is generally referred to as miRNA clusters. Even though many miRNA clusters have been greatly studied, most attention has been usually focused on functional impacts of clustered miRNA co-expression. However, there is also another compelling aspect about these miRNA clusters, their polycistronic nature. Being transcribed on a single RNA precursor, polycistronic miRNAs benefit from common transcriptional regulation allowing their coordinated expression. And yet, numerous reports have revealed striking discrepancies in the accumulation of mature miRNAs produced from the same cluster. Indeed, the larger polycistronic transcripts can act as platforms providing unforeseen post-transcriptional regulatory mechanisms controlling individual miRNA processing, thus leading to differential miRNA expression, and sometimes even challenging the general assumption that polycistronic miRNAs are co-expressed. In this review, we aim to address the current knowledge about how miRNA polycistrons are post-transcriptionally regulated. In

development of diseases have been extensively studied and recognized (Bartel, 2018;Kabekkodu et al., 2018;Rufino-Palomares et al., 2013). Yet, we are still discovering new mechanisms that post-transcriptionally regulate the accumulation of miRNAs themselves, which in turn directly influence the activity of each of these important regulatory molecules.
In animals, the canonical miRNA biogenesis pathway (Figure 1) entitles the sequential maturation of a long RNA Polymerase (Pol) II primary transcript (pri-miRNA) by two type III ribonucleases. In the first step, the pri-miRNA folds into a stem-loop structure, which is recognized and cleaved by the Microprocessor complex composed of the RNase III Drosha and a dimer of Pasha/DGCR8 (DiGeorge syndrome critical region 8), to generate the precursor miRNA (pre-miRNA). Then, the pre-miRNA is exported to the cytoplasm by the Exportin 5 factor and processed by another RNase III, called Dicer. The resulting 20-24 nucleotide (nt) long RNA duplex is loaded into one of the Argonaute (AGO) proteins and after unwinding the mature miRNA remains associated with AGO to form the effector RNA-induced silencing complex (RISC). The miRNA acts as a guide for the AGO protein to recognize its target RNAs. Upon base-pairing between the miRNA seed sequence (i.e. nt 2 to 7 or 8) and a complementary mRNA target sequence, RISC will recruit additional factors to either destabilize the target mRNA or impede its translation (see Bartel, 2018 for an extensive review on miRNA biogenesis and mode of action). Each step of this pathway is tightly controlled by a wide range of mechanisms, from regulating the transcription of the pri-miRNA to triggering mature miRNA decay, that ultimately fine-tune the accumulation of individual mature miRNAs (Treiber et al., 2019).
Since the first miRNA genes were discovered in animal genomes, researchers noticed their propensity to be localized in proximity and form groups of 2 or more tandemly arranged miRNAs, henceforth referred to as miRNA clusters. This observation was instrumental in helping in the systematic discovery of novel miRNAs. Based on the hypothesis that clustering is frequent and that novel miRNA genes can be located in the vicinity of previously known miRNAs, scanning of proximal genomic regions allowed to increase the miRNA repertoire (Altuvia, 2005;Sewer et al., 2005).
Another in silico approach aimed to predict miRNA clusters with no need for a pre-existing miRNA in their proximity. The main prerequisite for this algorithm was the clustered feature of paralogous miRNA genes (Mathelier & Carbone, 2013). With the annotation of miRNA genes in various animal species nearing completion, or already final as in human (Fromm et al., 2022), it appears that the original observations can be confirmed and that genomic clustering of miRNA genes is indeed higher than expected at random (Chaulk et al., 2016;Guo et al., 2014). According to Wang and colleagues who did an exhaustive analysis of miRNA gene distribution in six animal models, the proportion of miRNAs organized in clusters ranges from 17.3% in chicken to 52% in zebrafish (Wang, Luo, et al., 2016).
A miRNA cluster is usually defined by the presence on the same genomic locus of adjacent pre-miRNA sequences that are transcribed in the same orientation and are not separated by another transcription unit. Most of the clusters are small, ranging from 2 to 6 precursors, while some large clusters may contain dozens of miRNA genes such as the chromosome 19 miRNA cluster (C19MC) in human, miR-430 cluster in zebrafish and Sfmb2 miRNA cluster in mouse, which encompass 46, 58 and 72 pre-miRNAs, respectively (Malnou et al., 2019;Thatcher et al., 2008). Given their genomic proximity, they are believed to be transcribed as polycistrons (Baskerville & Bartel, 2005;Ozsolak et al., 2008;Saini et al., 2007) with one pri-miRNA carrying all the pre-miRNAs. This indeed makes sense given that polycistronic transcription allows several miRNAs to benefit from the same regulation by transcription factors and their co-expression allows to control several target genes at once (Lataniotis et al., 2017;O'Donnell et al., 2005;Segura et al., 2017;Sylvestre et al., 2007;Wee et al., 2012). Along the same line, several groups have shown that clustered miRNAs tend to target genes involved in the same biological process or pathway (Kim et al., 2009;Li et al., 2021;Wang, Luo, et al., 2016). This allows for a simultaneous and cooperative miRNA activity and may induce stronger response at the cellular level. Furthermore, many clusters contain miRNAs from the same miRNA family (with the same seed sequence), which share a set of common targets. Their simultaneous expression would therefore have an additive effect and potentiate a more robust target downregulation. Thus, polycistronic transcription has advantages that resemble the ones found in prokaryotic operon systems. Finally, it has been estimated that about 50% of conserved vertebrate miRNAs are located within miRNA clusters thereby indicating an evolutionary pressure to maintain these genes and their organization (Altuvia, 2005;Wang, Luo, et al., 2016).
Interestingly, although there is coordinated transcription of miRNAs within a cluster, many reports point to the fact that the mature miRNAs originating from this transcription unit can be expressed at very different levels and can follow different expression patterns in response to physiological stimuli (Contrant et al., 2014;Edwards et al., 2008;Gu et al., 2013;Hooykaas et al., 2016;Kotaki et al., 2017;Pratt et al., 2009;Ryazansky et al., 2011;Sempere et al., 2004;Welten et al., 2014;Yu et al., 2006;Zhou et al., 2018). This clearly indicates that there are regulatory mechanisms put in place at the post-transcriptional level to achieve differential expression of individual cluster members. Indeed, it has been shown that the same miRNA can be expressed to strikingly different levels if expressed individually or within a cluster (Chakraborty et al., 2012;Contrant et al., 2014). What is more, the relative expression can be also impacted by the pre-miRNA position within the polycistron. This indicates that there are features within a polycistronic pri-miRNA transcript that directly influence the coordinated processing of individual miRNAs. This in turn allows to achieve expression patterns perfectly tailored for their function. Hence, the long primary transcript can for example serve as a higher-order scaffold providing structural and sequence elements for auto-regulatory intramolecular interactions and/or help in the recruitment of external accessory factors. In addition, increasing evidence shows that within one cluster, individual pre-miRNAs do not behave as independent units, but show unexpected relationships making the whole cluster a dynamic and interdependent system.
In this review, we will focus on polycistronic miRNAs expressed in animals and in some of their viruses, since the latter strictly depend on the host cell machinery to express their miRNAs. We aim to summarize the current evidence of the post-transcriptional molecular mechanisms orchestrating the differential processing of pre-miRNAs within polycistronic primary transcript. Eventually, we hope to provide the reader with a comprehensive view of these events, which are crucial for the accumulation of clustered miRNAs at the right level.

| IMPORTANCE OF SEQUENCE, SECONDARY STRUCTURE, AND CLUSTER ASSISTANCE
Although there are examples of noncanonical, Drosha-independent, mechanisms of miRNA processing (e.g. mirtrons [Ruby et al., 2007]), polycistronic miRNAs do rely on the Microprocessor complex for separating the cluster members. It is well recognized that local secondary structure and primary sequence motifs present on the precursor hairpins play essential roles for efficient recognition by Drosha/DGCR8 (reviewed in Creugny et al., 2018;Michlewski & C aceres, 2019). The presence of several of these features has an additive effect, while their absence can prevent a given pre-miRNA processing (Fang & Bartel, 2015). Among the structural requirements, we can cite the length of the stem, the presence of a bulge in the stem and the size of the apical loop. Sequence-wise, a UG motif at the base of the stemloop, a UGU/GUG motif in the apical loop and a CNNC motif downstream of the pre-miRNA were all shown to positively contribute to the miRNA biogenesis ( Figure 2a). Therefore, various combinations of these pre-requisites on the different pre-miRNAs within a polycistron can influence the efficiency of their respective processing and account for their differential expression. Indeed, after resolving the secondary structure of a long miRNA cluster (mir-K10/12) expressed by the Kaposi's sarcoma associated herpesvirus (KSHV) by SHAPE (Selective 2 0 Hydroxyl Acylation) analyzed by Primer Extension analysis, Contrant et al. examined the features of the 10 miRNA precursors within the cluster and concluded that some of the hairpins are more prone to be processed than others. Importantly, a good correlation was found between the presence of the predefined characteristics cited above and the relative expression levels of the individual miRNAs (Contrant et al., 2014).
Interestingly, suboptimal hairpins have been shown to be enriched within polycistrons (Hutter et al., 2020;Shang et al., 2020). Given their consistent accumulation, this led to the hypothesis that since they are not ideal substrates for the Microprocessor, there must be other mechanisms to specifically boost their processing in an indirect manner.
F I G U R E 2 Importance of primary sequence and cis-acting elements for pri-miRNA cleavage. (a) Hypothetical miRNA hairpin presenting the sequence motifs (highlighted in yellow) and structural features known to enhance Microprocessor recognition/activity. Drosha and dicer cleavage sites are represented by green and orange triangles respectively. (b) Examples of miRNA pairs involved in cluster assistance. The main characteristics optimizing recognition of the helper hairpins or hindering the processing of recipient hairpins are featured. Structures are derived from available SHAPE data (Shang et al., 2020) or predicted with RNA structure (https://rna.urmc.rochester. edu/). (c) Principle of cluster assistance. Presence of the helper hairpin on the primary transcript enhances the processing of the recipient hairpin. (d) In some cases, the helper hairpin not only enables, but simultaneously limits the recipient expression (dashed arrow). Distance between the hairpins and other sequences such as poly(A) sites can also inhibit cluster assistance. (e) Schematic view of KSHV miRNA cluster with pri-miR-K1 and pri-miR-K3 having a positive regulatory effect on the remaining miRNAs within the cluster Focusing on bi-and tri-cistronic clusters in different animal species, several groups have noticed that deletion of one specific member of the cluster can severely impede the processing of another member, while deleting the other might have no effect (Fang & Bartel, 2020;Feederle et al., 2011;Haar et al., 2016;Hutter et al., 2020;Lataniotis et al., 2017;Shang et al., 2020;Truscott et al., 2016). A common feature of the precursors that fail to be processed in the absence of the other partner is that they present suboptimal characteristics and are therefore poor Microprocessor substrates per se. As such, they require the presence of a proximal helper precursor which meets canonical requirements for optimal Microprocessor processing. This phenomenon has been referred to as cluster assistance (Fang & Bartel, 2020;Hutter et al., 2020) (Figure 2b,c). Strengthening the evidence that optimal folding of the helper hairpin is key for cluster assistance, mutations designed to impact its recognition by the Microprocessor, such as disturbing proper base-pairing within the stem or reducing the size of the stem or of the apical loop, not only reduce the expression of the helper itself, but also of the recipient (Fang & Bartel, 2020;Hutter et al., 2020;Shang et al., 2020). Conversely, when the structure of the recipient hairpin is optimized, it can be processed on its own, even in the absence of its partner on the primary transcript (Fang & Bartel, 2020;Hutter et al., 2020;Shang et al., 2020;Truscott et al., 2016). In addition, cluster assistance is maintained in chimeric constructs where the helper is replaced by another heterologous miRNA stem-loop, unrelated in sequence, but with better hairpin characteristics (Fang & Bartel, 2020;Haar et al., 2016;Hutter et al., 2020;Shang et al., 2020;Truscott et al., 2016). Thus, the impact of the helper hairpin on its neighbor is not solely related to its sequence, structure or potential downstream function.
Other aspects of the phenomenon were also studied in detail. Namely, it was shown that cluster assistance does not operate directionally since swapping the order of helper and recipient stem-loops does not hinder it (Fang & Bartel, 2020;Hutter et al., 2020;Shang et al., 2020;Truscott et al., 2016). However, it seems to be mostly found in clusters where hairpins directly follow one another. Thus, when two suboptimal hairpins are placed immediately downstream of a helper pre-miRNA, there is a strong improvement in processing only for the proximal pre-miRNA while the distal one is only moderately impacted. However, placing the helper hairpin in between the two suboptimal ones provides efficient assistance to both of them (Shang et al., 2020). As to the maximal distance between the neighbor hairpins, it can be increased up to 1200 nt without lowering the potentiating effect (Hutter et al., 2020;Shang et al., 2020). It has been shown that interaction with RNA Polymerase II can be implicated in the enhancement of some pre-miRNA processing (Church et al., 2017). However, it does not seem to play a prominent role in cluster assistance since both RNA Polymerase II and III can synthesize the primary transcript without impacting the phenomenon (Fang & Bartel, 2020;Shang et al., 2020).
Deeper investigation of the molecular mechanism driving cluster assistance led to a model, in which the helper hairpin efficiently recruits the Microprocessor complex and then facilitates its transfer to the recipient stem-loop. The cleavage of the helper pre-miRNA might occur first, but is not necessarily required for the completion of the recipient cropping, since selective tethering of Microprocessor via the B-box system can mimic helper-driven assistance (Fang & Bartel, 2020). In addition, it has been proposed that the mechanism might act as an evolutionary driver of novel miRNAs emergence. Indeed, a recent miRNA, from an evolution point of view, might benefit from the presence of a nearby optimal pre-miRNA even though it has not yet acquired all the characteristics required for its own efficient processing. Taken together, these findings indicate that certain miRNA hairpins might possess additional function than to solely serve as precursors of mature miRNAs, which would include the stimulation of biogenesis within miRNA clusters.
However, not every case of cluster assistance fully complies with these principles. Thus, results obtained by the Delecluse and Frolov groups, studying the Epstein-Barr virus (EBV) mir-BHRF1~3 and the Drosophila mir-11~998 clusters respectively, indicate that the presence of optimal helper stem-loops can not only potentiate, but simultaneously limit aberrant overexpression of their neighbors ( Figure 2d). This would contribute to a more sophisticated fine-tuning effect relevant for physiological needs of each given mature miRNA (Haar et al., 2016;Truscott et al., 2016). Similarly, further sequences surrounding the hairpins or the presence of other elements in cis may weaken the stimulatory effect. For example, the presence of a downstream poly(A) signal limits the cluster assistance phenomenon, presumably due to competition between the Microprocessor and transcription termination machineries (Haar et al., 2016).
In the case of clusters containing more than two or three miRNAs, interdependence in processing between the hairpins reveals more complicated schemes. This has been shown by our group in the context of the viral cluster mir-K10/12 expressed by KSHV and containing 10 miRNA precursors. The presence in cis of two particular stem-loops, pre-miR-K1 and pre-miR-K3, within the cluster is necessary for efficient expression of the remaining miRNAs (Vilimova et al., 2021). They are both very efficiently processed by Drosha and the deletion of any of them globally impedes the entire cluster expression (Figure 2e). Moreover, insertion of a heterologous miRNA precursor in lieu of pre-miR-K1 efficiently rescues the expression, excluding the possibility that the sequence or the downstream function of this particular miRNA is required. While these observations are in adequation with cluster assistance in smaller clusters detailed above, in this particular case, the regulation seems to have broader implication. The stimulatory effect is neither reserved only to closest neighbors, nor to suboptimally folded hairpins. What is more, the local context on the transcript and the order of hairpins is of importance given that swapping the stem-loops impacts their processing efficiency (Contrant et al., 2014). While this is not excluded also for smaller clusters, the biogenesis of larger polycistronic pri-miRNAs likely relies on a synergy of mechanisms that may combine efficient Microprocessor recruitment by optimal hairpins, structural remodeling and implication of other factors in trans, as we will discuss below.

| INVOLVEMENT OF TRANS-ACTING FACTORS
In addition to cis regulatory elements, external factors interacting with the primary transcript also play an important role in modulating the processing of clustered miRNAs. Selective processing of specific pre-miRNAs within a cluster may rely on the activity of various RNA-binding proteins. These can either enhance the maturation by recruiting the Microprocessor or by remodeling suboptimal secondary structure, or on the contrary, compete with the cropping machinery or mask important recognition sites on the pri-miRNA. Numerous studies were dedicated to the discovery of such co-factors involved in miRNA biogenesis resulting in a collection of proteins known to be involved in the processing of specific miRNAs or more generally modulating the function of the Microprocessor itself. As an example, the protein hnRNPA1 has been shown to selectively bind to the terminal loop of miR-18a within the mir-17~92a cluster. This induces a relaxation of the stem-loop structure and facilitates the processing of miR-18a, while the other miRNAs within the cluster remain insensitive to hnRNPA1-mediated regulation (Guil & C aceres, 2007;Kooshapur et al., 2018;Michlewski et al., 2008). Another example comes from the studies of the 14q32/12F1 miRNA cluster, where binding of different proteins (Mef2a, HADHB, and CIRBP) to particular pre-miRNAs from the cluster seems to specifically modulate their conversion into mature miRNAs (Downie Ruiz Velasco et al., 2019;Welten et al., 2017).
Besides these co-factors, whose functions were recently reviewed elsewhere (Michlewski & C aceres, 2019), we would like to discuss a particular group of trans-acting factors, which were studied specifically in the context of miRNA polycistrons. More particularly, some proteins were shown to participate in the above-mentioned mechanism of cluster assistance. The Bartel and Herzog groups hypothesized that given the fact that the cleavage of the helper hairpin is not required, the mechanism is probably related either to the recruitment of the Microprocessor or of other factors that may help in transferring its activity on the recipient hairpin. When a purified Microprocessor was incubated in vitro with a primary transcript containing a helper and a recipient stem-loop, the latter failed to be efficiently processed. However, the phenomenon was functional if whole cell extract was used in the assay (Fang & Bartel, 2020). This demonstrated the need of additional factors for the assistance to occur. Functional screens allowed to identify two proteins, ERH (enhancer of rudimentary homolog) and SAFB2 (scaffold attachment factor B2), whose depletion selectively impacts the processing of recipient hairpins. What is more, both proteins directly bind to the Microprocessor (Fang & Bartel, 2020;Hutter et al., 2020;Kwon et al., 2020) and can interact with each other (Drakouli et al., 2017). In addition, both proteins are known to dimerize and dimerization of SAFB2 is likely to be necessary for efficient cluster assistance (Arai et al., 2005;Hutter et al., 2020). This led to a model in which these proteins mediate association of two or more Microprocessors, thereby allowing the simultaneous recognition and cleavage of neighboring hairpins (Figure 3a).
Protein co-factors are not the only trans-acting regulators of polycistronic miRNA maturation. Several reports have shed light also on the implication of long noncoding (lnc)RNAs. Through partial complementarity with regions on the primary transcript they can either inhibit correct folding, mask important sequences or help reposition the RNA scaffold for better recognition and processing by the Microprocessor, similarly to snRNAs during splicing. Among these lncRNAs are the so-called Transcribed ultra-conserved regions (T-UCRs) which are perfectly conserved between mouse, rat and human genomes (Bejerano et al., 2004). Such striking evolutionary retention suggests that they may play key functions in physiological or stress conditions. Liz et al. have shown that the T-UCR uc.283+A can directly bind to a complementary sequence in the lower stem of the miR-195 hairpin and selectively impede its recognition by the Microprocessor while the expression of the second miRNA within the cluster, miR-497, is not impacted (Liz et al., 2014). Interestingly, two other T-UCRs, uc.372 and uc.173, were shown to inhibit the processing of miR-195 (J. Guo et al., 2018;Xiao et al., 2018) (Figure 3b). Unfortunately, the expression level of miR-497 was not measured in these two studies, making it difficult to conclude whether the inhibitory effect is unique to miR-195 or impacts the entire cluster. An opposite situation was described by Soler et al., they showed that the binding of uc.160 to complementary sequences at the basal junctions of four out of the five members of the mir-376c~376a-1 cluster lead to an increased level of their processing (Soler et al., 2021) (Figure 3c). Finally, a recent paper by Han et al. reported a novel function of a circular RNA, circLONP2, in the processing of miR-17. CircLONP2 helps to recruit the Microprocessor to the miRNA hairpin by simultaneously binding to its stem and to the RNA helicase DDX1, which in turn directly binds DGCR8 (Figure 3d). This then leads to miR-17 upregulation in colorectal carcinoma (Han et al., 2020). However, how this impacts the processing of the remaining five members of the mir-17~92a cluster was not investigated.

| TERTIARY STRUCTURE OF THE PRIMARY TRANSCRIPT
Tridimensional folding of RNA molecules is essential for their proper function. Since the primary transcript of miRNA clusters may span over several kb, it is easily conceivable that these long RNA molecules are prone to higher-order structural arrangements and intra-molecular interactions. Thereby, a specific 3D architecture might explain differences in the processing efficiency of individual miRNA hairpins depending on their position within the molecular scaffold. This was best demonstrated for the mir-17~92a cluster, which contains six miRNA precursors within~800 nt.
F I G U R E 3 Contribution of trans-acting factors for miRNA processing. (a) Model of cluster assistance through Microprocessor dimerization involving ERH and SAFB2. (b,c) T-UCRs can bind to pri-miRNAs through complementary sequences highlighted in different colors. Inhibitory or stimulatory effect of this interaction on pri-miRNA processing is indicated by red and green arrows, respectively. Mature miRNA sequence is drawn in light blue. (d) circLONP2/DDX1-mediated recruitment of the Microprocessor to miR-17 hairpin Several biochemical and imaging studies have shown that the primary transcript of mir-17~92a folds into a compact globular structure in vitro (Chakraborty et al., 2012;Chaulk et al., 2011;Du et al., 2015). It was hypothesized that a tight tertiary structure could represent a barrier for Drosha cleavage. To test this, Chakraborty et al. split the cluster in two halves and rearranged their organization. This led to a loosen RNA structure and had an impact on the overall accumulation of the processing intermediates and mature miRNAs both in vitro and in vivo (Chakraborty et al., 2012). Interestingly, all the pre-miRNAs on the shuffled cluster were better processed indicating that in such relaxed conformation, they might be more accessible to the Microprocessor. The wild-type primary transcript would thus impose structural constraints to limit its processing and auto-regulate the amounts of produced miRNAs. Submitting the primary transcript to structural probing by SHAPE also revealed that nucleotides within certain apical loops and segments between the stem-loops are solvent-inaccessible indicating that they might be implicated in tertiary-structure contacts (Chakraborty & Krishnan, 2017). What is more, phylogenic comparisons of these regions have shown that they are well conserved supporting their functional importance for the maintenance of this compact conformation (Chakraborty et al., 2012). Similar conclusions were made by Chaulk et al., who also studied the impact of tertiary structure on mir-17~92a processing. Using single particle electron microscopy, they not only confirmed the compact folding of the primary transcript in a globular structure, but also noticed the presence of a dense core consisting of the 3 0 region of the cluster (Chaulk et al., 2011). Buried within this core by the 5 0 extremity of the transcript, miR-19b and miR-92a hairpins are less exposed to the Microprocessor complex. Indeed, upon deletion of the 5 0 part of the pri-miRNA, the sequestrated hairpins can be processed with increased efficiency. This indicates that the core miRNAs can only be processed after the cleavage of the more exposed 5 0 end which would liberate the remaining miRNA precursors. However, shortly after being released, the core domain seems to be targeted by exonucleolytic degradation limiting the timeframe during which it can be processed (Chaulk et al., 2011). Together, these mechanisms might explain the lower expression levels of the 3 0 miRNAs compared to the 5 0 miRNAs seen in some cell lines. Further investigation of the internal core architecture has provided evidence that miR-19b interacts with a non-miRNA stem-loop located between miR-19b and miR-92a hairpins (Figure 4a). This interaction is important to maintain the compaction of the core domain, since its disruption specifically increases the expression of miR-92a (Chaulk et al., 2014). Thus, tertiary structure-based regulation of differential expression of the individual miRNAs within the mir-17~92a cluster is achieved by limiting the accessibility of certain miRNAs.
Another model to explain the auto-inhibitory structure found in the mir-17~92a cluster was proposed by Du et al. They identified two cis-regulatory complementary sequences within the primary transcript that can directly bind to F I G U R E 4 Role of the three-dimensional structure in the regulation of mir-17~92a processing. (a) miRNA stem-loops internalized within the 3 0 core (highlighted in gray) are less accessible to the microprocessing machinery. Tight structure of the internal core is maintained by the interaction between the stem of pre-miR-19b and the non-miRNA stem-loop (NMSL) (Chaulk et al., 2011(Chaulk et al., , 2014. (b) Binding of the repressive domain (RD) and its complementary sequence RD* within the primary transcript forms an auto-inhibitory structure which limits the processing of all the miRNAs except miR-92a. The structure is resolved upon the cleavage (red arrowhead) of the transcript by the endonuclease CPSF3 and ISY1 (Du et al., 2015). (c) Mir-17~92a progressive hierarchical biogenesis model (Donayo et al., 2019) each other. This higher order conformation hinders the processing of the miRNAs on the cluster, except for miR-92a, which is located outside of the structure and thus remains insensitive to this steric inhibition. Interestingly, the structure imposed by the regulatory sequences is resolved by the action of the endonuclease CPSF3, assisted by the ISY1 protein, upstream of the miR-17 stem-loop. The processing intermediate termed "pro-miRNA" is then processed with higher efficiency compared to the noncleaved molecule. This results in an increase in the expression of five out of the six miRNAs within the cluster. Interestingly, there is a direct correlation between the expression of these five miRNAs and of CPSF3 and ISY1 during differentiation of mouse embryonic stem cells, the experimental model used for this study (Du et al., 2015) (Figure 4b).
The mir-17~92a cluster has been coined an "oncomiR" owing to the fact that it plays important roles in cell cycle, proliferation and apoptosis (reviewed in Olive et al., 2013). In numerous cancers of hematopoietic origin, the cluster undergoes genomic amplification leading to its aberrant expression. Donayo et al. studied the differential maturation of the individual miRNAs in the cluster in this particular oncogenic context. In contrast with the previous reports, the authors observed an accumulation of several processing intermediates indicating that there is a hierarchy in the primary transcript maturation, which occurs in a well-defined series of consecutive processing events. According to their model, the pre-miRNAs are cropped progressively from both extremities, leaving miR-18a and miR-19b as the last processed precursors, which correlates with their lower expression levels (Figure 4c).
Although they are all focused on the same miRNA polycistron, the above-presented models are not often compatible, if not contradictory. They are however consistent with the idea that the differential processing of miRNAs in a cluster is regulated structurally and does not necessarily rely on a uniform biogenesis process. The various conformational and processing patterns observed might be attributed to different experimental systems, given that the miRNAs originating from the cluster are known to present a dynamic and variable relative accumulation levels as a function of both developmental stage and tissue-type (Abasi et al., 2017). Moreover, different mechanisms controlling the maturation might be rewired between physiological and pathological conditions and most probably depend on the presence of potential co-factors that might be required to unlock a particular structural conformation.

| SPLICING-DEPENDENT REGULATION OF INTRONIC CLUSTERS
Like unique miRNAs, miRNA clusters can be intergenic, when their transcription is driven by their own promoter, but they can also be included into other transcriptional units. Although there are cases where miRNAs are harbored within exons or UTRs, it is more common to find them within introns of both coding and noncoding host genes (Chang et al., 2015;Kim & Kim, 2007;Rodriguez et al., 2004). Splicing is a major event in the processing of RNA Pol II transcripts and constitutes the first level of post-transcriptional regulation occurring upstream (or concomitantly) of pre-miRNA cleavage by the Microprocessor. If one host gene harbors several miRNAs, and they are not located within the same intronic or exonic unit, they cannot be considered as clusters. This is because splicing is thought to be completed during pre-mRNA synthesis and its fast dynamics makes it unlikely to find miRNAs from distant introns or exons physically linked on a single unprocessed RNA molecule. On the contrary, they are rapidly separated and follow distinct fates within their respective introns or exons. Even though long-distance interactions cannot be excluded, this does not leave much room for further regulatory mechanisms related to polycistronic organization.
However, splicing, and more particularly alternative splicing patterns can constitute another layer of miRNA regulation, if they are located within the same or in adjacent introns/exons. A detailed study of splicing patterns of the BamHI A Rightward Transcript (BART), which contains most of the miRNAs expressed by the EBV, revealed the importance of intron/exon retention for processing of the various miRNA precursors located on the long RNA molecule (Edwards et al., 2008). The 21 BART miRNA hairpins form two groups usually considered as 2 distinct clusters, lying 4.5 kb apart, but starting as one single transcript. Analysis of splice isoforms derived from this region in different EBV-infected cell lines showed that the BART miRNAs are grouped on four distinct introns (Figure 5a). Differential intron/exon retention between different cell lines correlates with variations in relative miRNA expression. This led to the hypothesis that distinct splicing patterns may account for the differential miRNA accumulation, by either increasing the efficiency of processing of some precursors or impairing the processing of others. This might be related to the assembly of spliceosomes and the splicing process itself at different splice sites. Indeed, it has been shown that the splicing machinery can interact with and have both stimulatory or inhibitory/competitive functions on the Microprocessor activity (Agranat-Tamir et al., 2014;Janas et al., 2011;Kataoka et al., 2009;Kim & Kim, 2007;Mattioli et al., 2014;Morlando et al., 2008).
Another interesting insight into the interplay between alternative splicing and pre-miRNA processing was revealed by the study of another viral miRNA cluster, mir-M8/M10 encoded by the Marek's disease herpesvirus (MDV-1). This cluster contains four miRNA precursors located within the first intron of the LAT noncoding RNA. Rasschaert et al. have discovered an alternative 3 0 splice site, which splits the cluster in two parts (Rasschaert et al., 2016). While the first two miR-M8 and -M6 remain intronic, the two other miR-M7 and -M10 become a part of the following exon. Interestingly, upon this splicing event, miR-M6, which is located just upstream of the novel splice site is not processed anymore by the Microprocessor. Instead, it is processed as a 5 0 tailed mirtron, whose 3 0 end is directly defined by splicing and 5 0 end is released by Drosha cleavage of the flanking pre-miRNA and the action of an unknown exoribonuclease. Mutation of the alternative splice site negatively impacts miR-M6 expression indicating that this event is required for its optimal expression. On the contrary, the alternative splicing itself is limited by the presence of miR-M7 stem-loop, which may help to balance the switch between the two processing pathways (Figure 5b). Regarding the alternative exon that is generated, it could serve two purposes. It can either cleave by the Microprocessor to give rise to pre-miR-M7 and -M10 or it can serve as a functional noncoding gene product and increase viral transcriptome diversity.
Mammalian genes also undergo alternative splicing at loci containing miRNA clusters. One interesting example is the mouse Mirg gene, which gives rise to a long noncoding RNA enriched in brain tissue. Melamed et al. found that the splicing pattern of this RNA varies during development, leading to progressive inclusion of an alternative exon in adult brain, as compared to embryonic brain tissue (Melamed et al., 2013). This exon is found inside an intron containing four miRNA precursors, miR-541, miR-409, miR-412, and miR-369. Interestingly, miR-412 hairpin overlaps with the alternative 3 0 splice site so that the alternative splicing and miR-412 processing are mutually exclusive. Therefore, this particular splicing event not only splits the miRNA cluster apart (which does not seem to impact the expression of the remaining miRNAs), but it specifically hinders the processing of miR-412 (Melamed et al., 2013) (Figure 5c). A slightly different situation has been described in the case of two other intronic clusters, mir-23b~24-1 and mir-106b~25. Both clusters contain three miRNAs each and alternative splice sites in between them can give rise to bicistronic subclusters, while the remaining miRNA finds itself into an exon (Agranat-Tamir et al., 2014;Ramalingam et al., 2014). At least in the case of miR-25, its exonic location negatively impacts its expression (Agranat-Tamir T A B L E 1 Experimentally validated post-transcriptional regulatory mechanisms in miRNA polycistrons

| CONCLUSION
The propensity of miRNA genes for clustering in operon-like structures has been known since the first miRNAs were discovered in animal genomes. However, we only start to understand that the functional significance of such polycistrons not only rely in their synchronized expression, but also in the emergence of previously unrecognized regulation modes. As we have seen above, various post-transcriptional molecular mechanisms can explain the differential accumulation of polycistronic miRNAs. Both cis-acting sequences or structural elements and trans-acting factors, such as RNAbinding proteins and noncoding RNAs, can recruit the Microprocessor complex, regulate the access of the miRNA processing machinery or optimize the processing of some pre-miRNAs. In addition, particular splicing patterns may be utilized to uncouple the processing of the precursors, thus adapting to conditions requiring changes in abundance of the different mature miRNAs. Table 1 provides a summary of miRNA clusters for which the mechanism at play in their post-transcriptional regulation was experimentally validated. Post-transcriptional regulation of miRNA clusters presents new challenges and avenues of research concerning miRNAs and indicates that we should consider them as important entities that are not only here to carry individual regulatory units. We also need to reconsider what defines a miRNA cluster per se. They are frequently defined based on genomic proximity and occasionally inferred from expression data clustering together miRNAs from a genomic region being co-expressed under certain conditions (Chaulk et al., 2016). However, as we have just seen, co-expression is not necessarily an imperative. What is more, identification of genomic miRNA clusters strongly depends on the selected threshold distance separating two miRNA loci. As an illustration, estimating the number of miRNA clusters within the human genome resulted in 99 distinct clusters encompassing 352 miRNA genes if the threshold used was 10 kb (Wang, Luo, et al., 2016), and 153 clusters containing 468 miRNAs when the cut-off was shifted to 50 kb (Kabekkodu et al., 2018). This criterium is quite arbitrary and does not provide information about the true polycistronic nature of the presumed miRNA clusters. It is indeed more challenging to evaluate the exact number of genuine polycistrons unless their primary transcripts are thoroughly characterized. Some efforts have been made to predict the promoter regions and transcription start sites (TSSs) or poly(A) signals for individual miRNA genes, which can be helpful for defining whether neighboring miRNA loci are or are not transcribed together as polycistrons (Chang et al., 2015;Chien et al., 2011;de Rie et al., 2017;Perdikopanis et al., 2021;Rodriguez et al., 2004;Saini et al., 2007;Yu et al., 2006). However, experimental validation is still rather scarce and concerns only a handful of well-expressed and studied clusters (Chaulk et al., 2016). The task may prove even more challenging if we take into account the cell-type and context-dependent expression and the fact that independent TSSs can give rise to subclusters under certain circumstances (Nayak et al., 2018;Ozsolak et al., 2008). Needless to say, future studies should explicitly differentiate between the two notions of genomic and polycistronic miRNA clusters, and deal with them carefully with respect to the scope of their research. Along the same line, a large majority of studies focusing on miRNA function and biogenesis still focus on single isolated miRNAs or their precursors, which brings about questions regarding the fate of their neighbors if they are located within a cluster. Knock-out experiments aiming to delineate the function of a particular miRNA may lead to skewed conclusions if the context of cluster is not properly addressed. Given the novel paradigms, some of the published results may need to be revised accordingly.
On another note, understanding the regulation at play within miRNA clusters opens new possibilities for manipulation and adaptation of novel genetic tools. As such, the design of new strategies for therapeutic intervention is particularly appealing. For example, one may take advantage of a well-designed single molecule targeting the primary transcript to impact the expression of several miRNAs at once. An RNA aptamer and an antisense oligonucleotide were proposed to inhibit the expression of the entire mir-17~92a and KSHV mir-K10/12 clusters, both involved in pathology (Lünse et al., 2010;Subramanian et al., 2015;Vilimova et al., 2021). Alternatively, the construction of artificial clusters expressing several therapeutical or synthetic miRNAs, each at a different level, could help to broadly modulate physiological or pathological signaling (Wang, Xie, et al., 2016;Yang et al., 2013). However, these approaches are still in their infancy and will require more work before they can be applied to real-life settings. Our knowledge of miRNA cluster regulation is still incomplete and fragmented. It is therefore problematic to define common or distinct features for each type of regulation and it is impossible to extrapolate the already available information to other clusters. To further complicate the picture, the regulation might be not only cluster-but also context-dependent, intertwined with the activity of specific factors and regulatory networks, that might be cell line/tissue-specific or rewired in diseased or stress conditions. In addition, while it is determinant, processing of the primary transcript is not the only level at which miRNAs are regulated. Indeed, the mature miRNA abundance results from a balance between efficient production and pre-and mature miRNA decay. Ultimately, all these aspects will have to be considered, if we want our knowledge of the regulation mechanisms during the processing of natural miRNA clusters to be translated one day into successful applications.

ACKNOWLEDGMENTS
We would like to thank Erika Girardi for critical reading of the manuscript and members of the laboratory for discussion.

FUNDING INFORMATION
Our work is funded by Agence Nationale de la Recherche through the Interdisciplinary Thematic Institute IMCBio, part of the ITI 2021-2028 program of the University of Strasbourg, CNRS and Inserm (ANR-10-IDEX-0002 and ANR-17-EURE-0023). MV is funded by a fellowship from the Ministère de l'enseignement supérieur, de la recherche et de l'innovation and by the Fondation ARC pour la recherche sur le cancer.