Separable roles for Microprocessor and its cofactors, ERH and SAFB1/2, during microRNA cluster assistance

In this study, Shang et al. describe how miRNA transcripts with suboptimal attributes are processed efficiently into small RNAs by the Microprocessor complex through a mechanism of miRNA cluster assistance. They show that Microprocessor cofactors like ERH and SAFB2 play distinct but critical roles in miRNA hairpin recognition and binding and Microprocessor transfer, constituting a complex regulatory pathway in clustered miRNA biogenesis.

Abstract

While most conserved microRNA (miRNA) transcripts harbor a suite of features that mediate their efficient biogenesis into small RNAs, some loci bear suboptimal attributes that enable additional layers of processing regulation. A notable example is cluster assistance, whereby a miRNA hairpin with suboptimal nuclear biogenesis can be enhanced by an optimal neighbor. This process involves local transfer of the Microprocessor complex, composed of the RNase III enzyme Drosha and its partner, DGCR8, in concert with cofactors such as ERH and SAFB1/2. However, the mechanisms that underlie miRNA cluster assistance remain largely unclear. Here, we gained insights into this process by integrating mutant cells of Microprocessor and its cofactors with analysis of miRNA structure–function variants, biochemical tests, and genome-wide profiling. We defined features of suboptimal miRNAs that render them dependent on cluster assistance and distinguished among a network of proposed interactions among Microprocessor and its cofactors to reveal a subset that is critical for cluster assistance. Most importantly, we used epistatic tests to separate and order the functional requirements for ERH and SAFB1/2 into a pathway. Our data indicate that ERH may engage in the process of Microprocessor transfer between hairpins, while SAFB factors (especially SAFB2) mediate recognition and stable binding of a suboptimal miRNA hairpin after Microprocessor transfer. Finally, we show how cluster assistance integrates into a feedback regulatory loop on Microprocessor via Drosha-mediated cleavage of a suboptimal miRNA hairpin in the DGCR8 transcript. Altogether, our findings reveal complex regulatory transactions during biogenesis of clustered miRNAs.

Since the discovery of the first microRNA (miRNA), lin-4, and its critical regulatory target, lin-14, by Ambros and Ruvkun in 1993 (Lee et al. 1993; Wightman et al. 1993), a large community of researchers continues to study how miRNAs are generated, regulate gene expression, and impact biology. The bulk of well-expressed miRNAs in animals is generated by a canonical biogenesis pathway. In short, following the transcription of primary miRNAs (pri-miRNAs), the nuclear Microprocessor complex (composed of the RNase III enzyme Drosha and its partner, DGCR8) recognizes and cleaves the hairpins to liberate the pre-miRNAs. Then, pre-miRNA hairpins are exported to the cytoplasm and cleaved by the RNase III enzyme Dicer, yielding ∼22 nt miRNA duplexes. These duplexes are loaded into Argonaute proteins, and a single strand is preferentially retained in the mature complex, which guides it to appropriate targets (Bartel 2018; Shang et al. 2023). As animal miRNA targets require only limited complementarity (∼7 nt) to positions 2–8 of the miRNA (Lai and Posakony 1997; Lai et al. 1998; Lai 2002; Brennecke et al. 2005), referred to as the seed region (Lewis et al. 2003; Bartel 2018), miRNAs tend to mediate extensive gene regulatory networks in metazoan species (McGeary et al. 2019).

As the gatekeeper for the canonical miRNA pathway, the Microprocessor complex must specifically select desired pri-miRNA substrates for efficient processing while rejecting a sea of other plausible RNA hairpins. Biochemical and genomic studies reveal that optimal pri-miRNA hairpins harbor single-stranded flanking regions, a double-stranded stem of ∼35 bp, a single-stranded terminal loop of >10 nt, and one or more sequence motifs within or close to the pri-miRNA hairpins (Zeng and Cullen 2005; Zeng et al. 2005; Han et al. 2006; Zhang and Zeng 2010; Auyeung et al. 2013; Ma et al. 2013; Fang and Bartel 2015; Nguyen et al. 2015; Kwon et al. 2016, 2019). Recent cryo-EM structural studies of Microprocessor and its substrates provide insights into how such features are recognized (Jin et al. 2020; Partin et al. 2020; Garg et al. 2024), though the full picture of their interactions is still unclear.

Although one might imagine that conserved miRNAs have been honed by evolution to acquire optimal biogenesis features, rarely does any biochemical or biological process run at full steam all of the time. For example, most canonical optimal miRNA hairpins harbor only a subset of the “menu” of positive biogenesis features (Fang and Bartel 2015; Kwon et al. 2019; Lee et al. 2023), which potentially enables regulation of processing. Perhaps more unexpected are miRNA loci that explicitly violate attributes of core canonical miRNA. In some cases, such highly suboptimal miRNA structures permit such substrates to escape the supervision of Microprocessor and exploit other ribonucleolytic strategies to access the miRNA pathway. For example, mirtron hairpins lack optimal dsRNA stems and ssRNA flanking regions but are instead cleaved by the splicing machinery (Okamura et al. 2007; Ruby et al. 2007). Transcription start site miRNA hairpins, exemplified by the mir-320 family, lack optimal dsRNA stems and terminal loops, but their short hairpins are directly transcribed by RNA polymerase II prior to Dicer processing (Xie et al. 2013; Zamudio et al. 2014). Additional RNA processing strategies can yield functional miRNAs via Drosha- or Dicer-independent mechanisms (Yang and Lai 2011), and in fact it is possible to efficiently generate mammalian miRNAs independently of RNase III enzymes (Maurin et al. 2012).

Unexpectedly, the processing of some suboptimal miRNA hairpins still depends on Microprocessor, including Dicer-independent mir-451 that bears a short stem and small terminal loop (Cheloufi et al. 2010; Cifuentes et al. 2010; Yang et al. 2010). How do such suboptimal hairpins access Microprocessor, when it is designed to exclude such substrates? Our group and others uncovered a mechanism termed cluster assistance that underlies the nuclear processing of Microprocessor-dependent suboptimal miRNA hairpins (Fang and Bartel 2020; Hutter et al. 2020; Shang et al. 2020). The critical insight is that intrinsically suboptimal nuclear biogenesis of a miRNA hairpin can be enhanced by an optimal miRNA hairpin within the same primary transcript. Mechanistic tests support that this involves ordered recruitment of Microprocessor to the optimal hairpin and local transfer to the suboptimal hairpin (Shang et al. 2020).

While the “recruit–transfer” model is focused on Microprocessor, several trans-acting factors also promote miRNA cluster assistance (Fang and Bartel 2020; Hutter et al. 2020). An elegant genetic screen took advantage of the mir-15a/16 cluster, where mir-16 bears optimal structure but mir-15a is suboptimal with a short lower stem. The use of dual miRNA reporters for both members enabled CRISPR/Cas9 screening to distinguish between hits that affect both miRNAs (including core factors in miRNA biogenesis and target regulation) and hits that might specifically aid suboptimal miRNA biogenesis (and thus score only with the miR-15a reporter). This revealed how scaffold attachment factors B1 and B2 (SAFB1/2) and enhancer of rudimentary homolog (ERH) selectively promote biogenesis of suboptimal miRNA hairpins (Hutter et al. 2020). Biochemical analysis of Microprocessor-associated proteins revealed that ERH binds directly to DGCR8 and participates in cluster assistance (Fang and Bartel 2020; Kwon et al. 2020).

Since this trio of initial studies, the DGCR8–ERH complex was solved (Kwon et al. 2020), and additional parameters of miRNA cluster assistance were enumerated (Shang and Lai 2023). Nevertheless, the mechanistic roles of these Microprocessor cofactors remain largely unclear. For example, it has not yet been distinguished whether they facilitate Microprocessor transfer between hairpins, stabilize or enhance cleavage activity of Microprocessor on suboptimal miRNA hairpins, or facilitate higher-order assemblies of Microprocessor. Moreover, in all these scenarios, it remains unclear whether ERH and SAFB1/2 work together in a complex with DGCR8/Drosha or whether there might be separable roles of these cofactors. Thus, current reviews of miRNA cluster assistance envision that numerous possibilities are plausible (Supplemental Fig. S1; Shang et al. 2023; Kim et al. 2025).

Further insights are complicated by the fact that both ERH and SAFB1/2 have numerous and still not fully defined roles beyond miRNA biogenesis. For example, ERH is one of the most highly interconnected factors among RNA binding protein (RBP) complexes (Street et al. 2024), and SAFB1/2 are paralogous proteins with dual DNA-binding and RNA-binding architecture (Norman et al. 2016). Accordingly, simple profiling of mutant conditions will inevitably reflect a complex mixture of direct regulatory effects, of which miRNA cluster assistance is only one, along with a suite of indirect consequences. Instead, we paired genetic mutants with a series of biochemical and genomic profiling experiments, along with panels of structure–function variants, which allowed us to specifically probe the requirements of Microprocessor cofactors and different steps of nuclear miRNA biogenesis. We found complex roles for these factors to support Microprocessor activity, including distinct effects on different types of suboptimal miRNAs, and separable roles for ERH and SAFB2 during Microprocessor recruitment and transfer.

Results

Both ERH and SAFB2 are required for efficient processing of suboptimal mir-451

To study the roles of ERH and SAFB1/2 in miRNA cluster assistance, we generated a collection of mutant HEK293T cells. While we succeeded in recovering protein-null SAFB2 mutants, following multiple rounds of engineering, we found that all ERH and SAFB1 mutant clones retained a wild-type allele (Supplemental Fig. S2). We conclude that SAFB1 and ERH are both cell-essential in HEK293T cells. Because multiple core miRNA factors can be fully deleted from HEK293T cells (Bogerd et al. 2014; Aguado et al. 2017), this indicates that ERH and SAFB1 must harbor non-miRNA functions. Accordingly, we further depleted these factors using ERH-RNAi in ERH^+/− cells and SAFB1-RNAi in SAFB1^+/−, SAFB2^−/− cells (Fig. 1A). We also made a triple mutant by knocking down ERH and SAFB1 in ERH^+/−; SAFB1^+/−, SAFB2^−/− cells. These represent the strongest individual and combined depletions that we could obtain for these cofactors. Nevertheless, a “silver lining” to these efforts was that cells were not adapted to complete knockout of these factors, which is an advantage to the study of these models.

Figure 1.

Biogenesis of suboptimal mir-451 requires cluster assistance from ERH and SAFB1/2. (A) Deletion and/or depletion of ERH and SAFB1/2 in HEK293T cells. ERH mutant cells were generated by shRNA-mediated knockdown in an ERH^+/− heterozygous cell line. SAFB1/2 double-mutant cells were generated by knockdown of SAFB1 in a SAFB1^+/−, SAFB2^−/− cell line. ERH+SAFB1/2 triple-mutant cells were generated by knockdown of ERH and SAFB1 in an ERH^+/−; SAFB1^+/−, SAFB2^−/− cell line. Western blot validates the strong loss of ERH and SAFB1 proteins and the absence of SAFB2 in the respective mutant genotypes. β-Tubulin was probed as a control. (B) Northern blotting to assay processing of the mir-144/451 cluster or solo mir-451 in wild-type or cofactor-depleted HEK293T cells. Cotransfected mir-375 and endogenous let-7a and U6 snRNAs were probed as controls. RNA size markers (in nucleotides) are shown at the left. Microprocessing of suboptimal pri-mir-451, but not optimal pri-mir-144 and pri-mir-375, is significantly impaired in ERH and/or SAFB1/2 mutant cells. However, lack of the neighboring optimal mir-144 hairpin ablated pri-mir-451 processing even with all the cofactors present. Normalized quantifications of (pre + mature) miR-451 total levels to U6 snRNA levels using Multi Gauge V3.0 are shown below the gel. (C) Rescue of pri-mir-451 processing in mutant cells. While ERH overexpression can restore mir-451 biogenesis in ERH mutant cells, only SAFB2 overexpression restored mir-451 biogenesis in SAFB1/2 and ERH+SAFB1/2 mutant cells.

We examined the biogenesis of a mir-144/451 construct transfected into these cells. To note, miR-451 is a noncanonical miRNA that transits an unusual biogenesis pathway, yielding an atypical banding pattern on Northern blots (Cheloufi et al. 2010; Cifuentes et al. 2010; Yang and Lai 2010; Yang et al. 2012). To summarize, cleavage of pri-mir-451 by Drosha/DGCR8 yields a 42 nt pre-mir-451 hairpin that is too short to be cleaved by Dicer and instead loads directly into Ago proteins. Of the four human Ago family members, only Ago2 is capable of cleaving its 3′ arm, yielding a 30 nt product. This is further 3′-trimmed by PARN (Yoda et al. 2013) into heterogeneous, active, miR-451 species, which differ from typical miRNA products that are typically a distinct species in the 21–24 nt range (Fig. 1B).

Extending previous work on the dependency of miR-451 on ERH (Fang and Bartel 2020; Kwon et al. 2020), we found that maturation of suboptimal mir-451 was substantially and selectively inhibited in ERH single-mutant, SAFB1/2 double-mutant, and ERH/SAFB1/2 triple-mutant cells (Fig. 1B). In contrast, mir-144 was unaffected by any of these mutations; cotransfected optimal mir-375 and endogenous let-7a were also unaffected (Fig. 1B). Notably, the nuclear processing of a solo mir-451 construct lacking its optimal mir-144 hairpin neighbor was strongly abrogated in wild-type cells and was not further repressed by loss of ERH and/or SAFB1/2 (Fig. 1B). We conclude that the presence of the optimal miRNA hairpin is the first and most essential step for cluster assistance and that cofactors opt into the following steps to enhance processing of the adjacent suboptimal hairpin.

SAFB1 and SAFB2 have similar domain organization and sequences. To evaluate the extent of their functional redundancy, we compared SAFB1-RNAi in a SAFB1^+/− cell line with SAFB2-null cells. Loss of SAFB1 did not affect mir-451 processing, while mutation of SAFB2 inhibited miR-451 biogenesis to an extent comparable with that of ERH depletion (Supplemental Fig. S3A). However, combined SAFB1/2 loss had a stronger effect than SAFB2 loss alone (Supplemental Fig. S3A). These data mirror results with the mir-15a/16 suboptimal miRNA cluster in murine Baf3 cells (Hutter et al. 2020). Thus, SAFB1/2 have overlapping roles in cluster assistance, but SAFB2 plays a dominant role in cell lines from different species. We emphasize that SAFB1 appears to be cell-essential in HEK293T, while protein-null SAFB2 lines were easily recovered (Fig. 1A). Accordingly, SAFB1/2 paralogs must have some distinct functions; that is, the preferred role of SAFB2 in cluster assistance is not solely due to, for example, its preferred expression in HEK293T cells.

Next, we conducted rescue assays in our panel of mutant cells by transfecting cognate and noncognate cluster assistance factors. Ectopic expression of all these factors rescued miR-451 biogenesis in their corresponding mutants, confirming that their defects were not due to unanticipated genetic aberrations. More interesting were the results from cross-rescue tests. Ectopic ERH could not rescue SAFB1/2 mutants, suggesting that SAFB proteins exert a role nonredundant from ERH (Supplemental Fig. S3B). In addition, only SAFB2, but not SAFB1, rescued mir-451 maturation in SAFB1/2 double-depleted cells (Fig. 1C). This is consistent with our data that SAFB2 is more important than SAFB1 for cluster assistance in HEK293T cells. Unexpectedly, SAFB2 partially rescued miR-451 processing in both ERH mutants (Supplemental Fig. S3B) as well as ERH/SAFB1/2 triple-mutant cells (Fig. 1C). Such epistasis data hint that not only do SAFB and ERH have distinct roles but that SAFB2 may function downstream from ERH. These findings informed our subsequent mechanistic tests.

Specific mutations of mir-451 decouple the biogenesis requirements for ERH and SAFB2

Meta-analysis of small RNA profiling showed that a substantially overlapping set of miRNAs depends on both ERH and SAFB1/2 (Kwon et al. 2020), suggesting that they function in a common process underlying cluster assistance. Such a scenario might be ruled out if any miRNA was explicitly shown to depend on only one of these factors, but none have yet been characterized. Indeed, our tests of suboptimal mir-451, previously shown to depend on ERH, reveal that it also depends on SAFB1/2 (Fig. 1).

Accordingly, we attempted to decouple their roles by introducing mutations into mir-451, building on our extensive engineering of the mir-144 and mir-451 hairpins (Shang et al. 2020, 2022; Shang and Lai 2023). We were guided by previous observations that the suboptimal mir-15a hairpin has a short lower stem, while mir-451 has a small terminal loop and a short upper stem. Thus, we designed mir-451 mutants (within the context of the mir-144/451 cluster) that lack one or more suboptimal hairpin features; namely a single-stranded apical loop of >10 nt, a double-stranded upper stem of ∼22 bp, and a stable lower stem of ∼11 bp (Fig. 2A,B). As a control, we optimized the pri-mir-451 hairpin (451LSM) to include all of the above optimal structural features, which bypasses the need for an optimal miRNA neighbor (Shang et al. 2020).

Figure 2.

Engineered suboptimal miRNAs separate dependency on ERH and SAFB2. (A) Schematic of mir-144/451 variants. We constructed an optimized pri-mir-451 (451LSM) or variants with different suboptimal structural features, such as a small apical loop, a short upper stem, an unstable lower stem, or combinations of these features. (B,C) Northern blotting for the processing of mir-144/451 variants with a stable lower stem (B) or an unstable lower stem (C) in WT, ERH-depleted, and/or SAFB2-depleted cells. (B) Compared with wild-type mir-451 containing both a small apical loop and a short upper stem, mir-451 variants with only one of the two suboptimal features (451SM or 451LM) were less dependent on ERH and SAFB2. (C) A mir-451 variant bearing only an unstable lower stem (451LSM-LSU) was especially dependent on SAFB factors but only mildly dependent on ERH. When mir-451 harbors an unstable lower stem and another suboptimal feature (451SM-LSU, 451LM-LSU, or 451-LSU variants), its processing was fully suppressed even when paired with an optimal mir-144 neighbor and in the presence of all cluster assistance factors. The quantified ratios were normalized to the red lane in each mir-451 mutant set.

We transfected mir-451 mutants into wild-type and cofactor-depleted cells and compared their biogenesis using Northern blotting. Consistent with our previous data, the optimized mir-451 hairpin (451LSM) was independent of ERH and SAFB1/2 (Fig. 2B). While the processing of mir-451 mutants with either a small apical loop (451SM) or a short upper stem (451LM) showed some dependence on ERH and SAFB2, their dependence was weaker than wild-type pri-mir-451, which harbors both a small apical loop and a short upper stem (Fig. 2B). Because the presence of the neighboring optimal mir-144 hairpin alone can significantly enhance the processing of pri-mir-451SM and pri-mir-451LM (Shang et al. 2020), the roles of ERH and SAFB2 in their processing are limited.

We next examined a mir-451 mutant bearing an optimized loop but harboring an unstable lower stem (451LSM-LSU). Strikingly, this mutant was only subtly dependent on ERH but strongly dependent on SAFB1/2 (Fig. 2C). This implies that SAFB2 is particularly critical when the miRNA hairpin lacks a stable lower stem. Importantly, the behavior of 451LSM-LSU showed that it is possible to decouple the requirements for the two types of cluster assistance factors. Finally, we mutated the mir-451 lower stem in combination with another suboptimal feature—either a small apical loop (451SM-LSU), a short upper stem (451LM-LSU), or both (451-LSU) (Fig. 2A). All of these variants fully blocked the biogenesis of miR-451 even in the presence of an optimal miRNA neighbor and with all endogenous cofactors intact (Fig. 2C). Altogether, our tests on various mir-451 mutants reveal that miRNAs with different suboptimal features exhibit selective dependence on ERH and SAFB2.

ERH promotes cluster assistance, but this process does not require DGCR8/ERH interaction

With these data and reagents in hand, we proceeded to dissect the roles of Microprocessor cofactors, beginning with ERH. Human ERH is only 104 amino acids long, lacks an RNA binding domain, and does not interact with RNA directly. However, it is a constituent of numerous RBP complexes (Street et al. 2024) and associates with Microprocessor via DGCR8 (Fig. 3A; Kavanaugh et al. 2015; Fang and Bartel 2020; Kwon et al. 2020). Thus, we tested for interactions between ERH and miRNA pathway factors using coimmunoprecipitation (co-IP) assays. ERH interacted robustly with DGCR8 and weakly with Drosha (Fig. 3B), consistent with prior studies. Co-IP assays with luciferase and Dicer served as negative controls.

Figure 3.

Interaction of ERH and DGCR8 is not required for cluster assistance. (A) Schematics of DGCR8 domains and protein/cofactor binding regions. (B) ERH interactions with miRNA pathway factors. Co-IP tests showed that ERH has strong interaction with DGCR8, moderate interaction with Drosha (likely indirect), and no interaction with Dicer or control firefly luciferase. (C) The dimerization ability of ERH is required for its interaction with DGCR8. All dimerization mutants (Y79A, Y79E, T78A/Y79A, and previously reported I5R/L7R) of ERH abrogated their interactions with DGCR8. (D) Northern blotting showed that an ERH dimerization mutant (I5R/L7R) exhibited less efficient restoration on suboptimal pri-mir-451 processing than wild-type ERH, indicating that its dimerization is required for cluster assistance. (E) ERH interactions with wild-type and mutated DGCR8 by co-IP assays. N-terminally truncated DGCR8[146–773] fails to interact with ERH, while DGCR8 dimerization mutants (P351A and C352A) maintain interactions with ERH. (F,G) Both wild-type DGCR8 (F) and DGCR8[146–773] (G) can homodimerize in both WT and Drosha-KO cells. Thus, ERH is not essential for the dimerization ability of DGCR8. (H) In a DGCR8-KO cell line (from the Shuo Gu laboratory), both full-length DGCR8 and DGCR8[146–773] rescued mir-144/451 processing. In fact, DGCR8[146–773] that cannot bind ERH provided stronger rescue for pri-mir-451 processing than WT DGCR8. Meanwhile, DGCR8 dimerization mutants cannot (C352A) or only partially (P351A) restore mir-144/451 processing. Thus, the interaction between ERH and DGCR8 is dispensable for cluster assistance. (I,J) Tethering assays for suboptimal mir-451 (I) or mir-374b (J) confirm that the ERH-interacting region in DGCR8 is not essential for cluster assistance.

Structural studies revealed how human ERH homodimerizes and interacts with the N termini of DGCR8 dimers, forming a stable 2:2 complex (Kwon et al. 2020). To assess whether ERH homodimerization is required for interaction with DGCR8, we tested a previously reported ERH mutant (I5R/L7R) and generated additional mutants (Y79A, Y79E, and T78A/Y79A) based on its structure (Kwon et al. 2020). We used pull-down assays to confirm that all of these lack homodimerization capacity (Supplemental Fig. S4A). These ERH dimerization mutants almost completely failed to associate with DGCR8, indicating that the interaction interface of these factors requires ERH dimerization (Fig. 3C). In a rescue assay, the ERH dimerization mutant (I5RL7R) did not enhance processing of the suboptimal mir-451 hairpin (Fig. 3D).

These results clearly confirmed interactions between ERH and DGCR8 and that ERH requires dimerization to mediate miRNA cluster assistance. However, can we conclude from this that ERH:DGCR8 interaction is essential for cluster assistance? To test this, we truncated the N terminus with DGCR8[146–773], which fully abrogates interaction with ERH (Fig. 3E), consistent with published data (Kwon et al. 2020). In addition, we tested DGCR8 mutants that compromise hemin binding (P351A) (Barr et al. 2011, 2012) or lack both hemin-binding and dimerization abilities (C352A) (Faller et al. 2007). Importantly, both mutants can still interact with ERH (Fig. 3E). Notably, truncated DGCR8[146–773] that cannot bind ERH still homodimerized as efficiently as WT DGCR8 in both wild-type and Drosha-null cells (Fig. 3F,G). Therefore, DGCR8 dimerization does not require DGCR8–ERH interaction or ERH dimerization.

Using these reagents, we assessed whether the ERH-interacting domain in DGCR8 is required for cluster assistance. We expressed wild-type and variant DGCR8 proteins in DGCR8-null cells (Dai et al. 2016), which accumulated to comparable levels (Supplemental Fig. S4B), and compared their activities on mir-144/451 processing. Surprisingly, DGCR8[146–773] completely rescued biogenesis of both miR-144 and miR-451 in DGCR8-KO cells and even exhibited higher activity than wild-type DGCR8 to support biogenesis of suboptimal mir-451 (Fig. 3H). On the other hand, loss of DGCR8 dimerization ability, especially the C352A mutant, abrogated the processing of optimal mir-144, suboptimal mir-451, and endogenous let-7a hairpins (Fig. 3H). This was the case regardless of whether or not the ERH-interacting domain was present. We repeated these tests in an independent DGCR8-KO cell line and obtained similar results (Supplemental Fig. S4C).

We further tested the dependence on ERH–DGCR8 interaction for cluster assistance using our established λN-BoxB tethering assay (Shang et al. 2020). Direct recruitment of Microprocessor (containing full-length or N-terminal-truncated DGCR8) to the neighboring miRNA hairpin enhanced suboptimal mir-451 processing (Fig. 3I). Notably, tethering ERH to the optimal miRNA hairpin also enhanced mir-451 biogenesis, but the dimerization and DGCR8-interacting mutants (ERHI5RL7R) could not do so (Fig. 3I). We obtained similar results upon tethering the panel of ERH variants near another suboptimal miRNA, mir-374b (Fig. 3J).

These tests support two key conclusions regarding cluster assistance. First, recruitment of Microprocessor to a miRNA cluster by tethering ERH via interactions of ERH:DGCR8 can effectively promote suboptimal miRNA processing. However, if a sufficient level of Microprocessor is locally present, then the interaction between ERH and DGCR8 is dispensable for functional cluster assistance.

Minimal SAFB2 promotes cluster assistance, but this process does not require Drosha/SAFB2 interaction

We next explored the mechanism of SAFB1/2 in miRNA cluster assistance. As SAFB1/2 are broadly involved in numerous cellular processes via multiple protein domains, we took advantage of SAFB2[512–726] (Fig. 4A), a minimal separation-of-function fragment that is sufficient to promote miRNA cluster assistance (Hutter et al. 2020). We first validated that truncated SAFB2 could rescue suboptimal mir-451 processing to an extent comparable with that of full-length SAFB2 in both SAFB2-KO and SAFB1/2 double-mutant cells (Fig. 4B,C). Optimal miRNAs tested were not affected by truncated SAFB2, confirming that this minimal fragment has specific and sufficient roles in suboptimal miRNA processing (Fig. 4B,C).

Figure 4.

The N-terminal region of Drosha is dispensable for cluster assistance. (A) Schematics of Drosha and SAFB2 domains and protein binding regions. (B,C) A minimal fragment of SAFB2 containing only the RE-rich domain and a short upstream region can replace full-length SAFB2 during suboptimal miRNA processing. In rescue assays of SAFB2 single-mutant (B) or SAFB1/2 double-mutant (C) HEK293T cells, SAFB2[512–726] restored nuclear cleavage of suboptimal pri-mir-451 as efficiently as full-length SAFB2. At the same time, ectopic SAFB1 and ERH were incapable of rescuing mir-451 biogenesis in SAFB2-KO cells. (D) Co-IP tests showed very weak interactions between Drosha and functionally active SAFB2[512–726] in either WT or Microprocessor-null cells. (E) Expression of full-length and N-terminal-truncated Drosha in Drosha-KO HEK293T cells. (F) Restoration of mir-144/451 and control mir-375 processing in Drosha-KO cells by both wild-type Drosha and N-terminal-truncated Drosha (391–1374). (G) Restoration of mir-144/451 processing in Drosha/DGCR8-dKO cells (from the Shuo Gu laboratory) by wild-type Drosha or N-terminal-truncated Drosha (391–1374) and DGCR8[146–773]. (H) DGCR8 can coimmunoprecipitate with functionally active SAFB2[512–726], and this interaction was substantially strengthened in Drosha-KO cells.

It was reported that SAFB2 and Drosha interact, with the minimal regions corresponding to SAFB2[512–726] and Drosha[1–390] (Fig. 4A; Hutter et al. 2020). However, our co-IP tests did not detect substantial interactions between functional SAFB2[512–726] and either full-length or N-terminally-truncated Drosha in either wild-type or Microprocessor knockout cells (Fig. 4D; Supplemental Fig. S5A). To further test whether the interaction of Drosha and SAFB2 is essential for cluster assistance, we performed rescue assays for suboptimal miRNA processing in Drosha-KO cells using full-length or N-terminally truncated Drosha lacking the reported SAFB2 binding region (Fig. 4E). Interestingly, Drosha[391–1374] restored the processing of optimal mir-144 and suboptimal mir-451 in Drosha-KO cells to levels equivalent to that of full-length Drosha (Fig. 4F). Accordingly, neither the N-terminal region of Drosha nor its potential interaction with SAFB2 is essential for suboptimal miRNA processing.

We performed more stringent tests by expressing N-terminal truncation variants of Drosha and DGCR8 in Drosha/DGCR8 double-knockout (dKO) cells. These cells cannot mature endogenous or exogenous canonical miRNAs and remain fully incompetent even when singly expressing Drosha or DGCR8 (Fig. 4G). Only upon coexpression of Drosha and DGCR8 did we observe normal maturation of ectopically expressed mir-144/451 and mir-375, along with endogenous let-7a and miR-16 (Fig. 4G). Coexpression of either N-terminal deletion variant of Drosha and DGCR8 (DroshaΔN or DGCR8ΔN), along with the reciprocal full-length partner, provided near-normal rescue. Remarkably, coexpression of DroshaΔN and DGCR8ΔN also substantially rescued Drosha/DGCR8-dKO cells, including for processing of suboptimal mir-451 (Fig. 4G). While the efficiency of miR-451 processing was slightly lower than with rescue by full-length Microprocessor, we found that other canonical miRNAs, both exogenous and endogenous, were similarly compromised (Fig. 4G). Therefore, dual N-terminally truncated Microprocessor does not function quite as well as wild type, but the combined loss of their ERH- and SAFB2-interacting domains does not specifically compromise cluster assistance.

Adding further complexity to the network of potential interactions among Microprocessor cofactors, ERH was reported to associate with the C-terminal RG/G-rich regions of SAFB1/2 (Drakouli et al. 2017). However, steady-state interactions between ERH and SAFB2[512–726] were weak at best in either wild-type or Microprocessor mutant cells (Supplemental Fig. S5B). Because SAFB2[512–726] is competent for cluster assistance but lacks the RG/G-rich region, the potential interaction between ERH and SAFB2 does not seem to be required for cluster assistance.

If SAFB2 does not mediate cluster assistance via the reported interactions with Drosha or ERH, how might it interface with Microprocessor? Unexpectedly, we found that SAFB2 could coimmunoprecipitate DGCR8 modestly but that this interaction was strongly enhanced in Drosha-KO cells (Fig. 4H). Although SAFB2 is not required for DGCR8 dimerization or for interaction of Drosha and DGCR8 (Supplemental Fig. S5C), this still provides a possible means for SAFB2 to recruit Microprocessor at suboptimal miRNAs.

Separable functions for ERH and SAFB2 during miRNA cluster assistance

We proposed a Microprocessor “recruit–transfer” model for cluster assistance involving the following: step 1: recruitment of Microprocessor and cleavage of the optimal miRNA hairpin, step 2: release and transport of Microprocessor to the suboptimal miRNA hairpin, and step 3: recognition and processing of the suboptimal miRNA hairpin by Microprocessor (Shang et al. 2020). This framework is agnostic to the roles of cofactors such as ERH and SAFB2, which could either assist one or more steps in this model or conceivably play orthogonal roles (Supplemental Fig. S1). We used tethering assays in our panel of mutant cells to dissect settings that do or do not require these cofactors (Fig. 5A).

Figure 5.

ERH and SAFB2 operate at distinct steps of cluster assistance. (A) Schematics of variant mir-451 expression constructs, some of which are modified to contain a BoxB element that recruits λN-DGCR8 and thus Microprocessor. (B) Direct tethering of either full-length λN-DGCR8 or N-terminal-truncated λN-DGCR8[146–773] to the neighboring miRNA hairpin in the cluster can efficiently enhance suboptimal mir-451 processing in WT cells. In this setting, both ERH and SAFB2 are needed for functional cluster assistance. (C) Tethering λN-DGCR8 to the solo mir-451BoxB hairpin also promotes its processing and now bypasses the requirement for ERH. However, SAFB factors are still required in this context. Tethered λN-DGCR8[146–773] showed similar genetic dependencies for solo mir-451 processing, further demonstrating that ERH is dispensable for recognition and processing of the suboptimal miRNA hairpin. (D) Overexpression of Microprocessor can enhance mir-451 processing in the absence of neighboring optimal mir-144. In this context, SAFB factors but not ERH are required for enhanced biogenesis of solo mir-451. This confirms that SAFB factors function after Microprocessor finds the suboptimal miRNA hairpin. (E) Model for sequential roles of ERH and SAFB2 during miRNA cluster assistance. In step 1, Microprocessor is recruited by the optimal miRNA hairpin for its processing, which does not involve the cofactors. In step 2, the transfer of the Microprocessor to the suboptimal miRNA hairpin is assisted by ERH. In step 3, the suboptimal miRNA hairpin is recognized and processed by Microprocessor with the help of SAFB2, potentially via interaction with DGCR8.

As shown in Figure 1B, the biogenesis of pre-mir-144 does not require ERH or SAFB2, reflecting that these factors do not impact the first step of Microprocessor recruitment and cleavage. We took advantage of this by directly tethering Microprocessor via λN-DGCR8 to mir-144 using a BoxB element (Fig. 5A). Nuclear cleavage of the mir-144BoxB hairpin was similar in wild-type and mutant cells (though this extension blocked subsequent processing by Dicer), while suboptimal mir-451 was efficiently processed only in wild-type cells (Fig. 5B). Curiously, recruiting Microprocessor with N-terminally truncated λN-DGCR8[146–773] still yielded the pre-mir-144BoxB hairpin regardless of the cofactors but only strongly enhanced mir-451 in the presence of both ERH and SAFB2 (Fig. 5B). This provided evidence that ERH and SAFB2 enhance suboptimal miRNA processing after Microprocessor recruitment and cleavage of the optimal miRNA hairpin and in a manner independent of ERH and DGCR8 interaction.

We next separated steps 2 and 3 in the “recruit–transfer” model by tethering Microprocessor to solo mir-451BoxB (Fig. 5A); that is, without neighboring mir-144 (Shang et al. 2020). With this substrate, we found that SAFB2, but not ERH, was required (Fig. 5C). Moreover, direct recruitment of λN-DGCR8[146–773] to mir-451 efficiently enhanced its processing, further indicating that interaction of ERH and DGCR8 can be bypassed during cluster assistance (Fig. 5C).

Finally, we coexpressed solo mir-451 with Microprocessor. In wild-type cells, this condition can dramatically enhance biogenesis of a suboptimal solo miRNA (Shang et al. 2020). In the presence of ectopic expression of Microprocessor in cofactor mutant cells, the biogenesis of miR-451 could bypass ERH but not SAFB2 (Fig. 5D). We conclude that ERH is not involved in the recognition and processing of the suboptimal miRNA hairpin in step 3 but instead mediates the movement of Microprocessor between the optimal and suboptimal hairpins in step 2. On the other hand, we infer that SAFB2 mainly promotes recognition and/or cleavage of the suboptimal miRNA hairpin in step 3, though it may also play a role during Microprocessor transport. Altogether, these tests clearly reveal separable and likely sequential roles for ERH and SAFB2 during miRNA cluster assistance (Fig. 5E).

Global effects of ERH and SAFB2 on miRNA expression

Previous studies of global effects of ERH and SAFB1/2 on endogenous miRNAs involved manipulations of different cell lines (HEK293T cells, K562 cells, hMSCs, and Ramos cells) (Fang and Bartel 2020; Hutter et al. 2020), which complicates their meta-analysis (Kwon et al. 2020). Moreover, mutant combinations of ERH with SAFB1/2 have not yet been assessed. To enable direct comparisons, we generated biologically replicate small RNA data sets across our panel of HEK293T mutant settings; namely, ERH single depletion, SAFB1/2 double depletion, and ERH+SAFB1/2 triple depletion. Importantly, to be able to assess biogenesis effects on newly synthesized miRNAs, we used SLAM-seq (Herzog et al. 2017; Reichholf et al. 2019) with 6 h labeling. As we sequenced these data deeply (∼60 million reads per library), we were able to broadly assess both steady-state and nascent populations of miRNAs (11 million to 15 million reads per library). We provide processing statistics and quality control analyses of these data sets in Supplemental Figure S6, A–G. Importantly, the replicates for each genotype were highly correlated for the abundance of individual miRNAs (Supplemental Fig. S6H). Detailed statistics of the libraries are in Supplemental Table S1.

As a preliminary check of our data, we performed principal component analysis (PCA) and found that the replicates were well correlated and that the genotypes were substantially distinct from each other (Fig. 6A). The SAFB1/2 and ERH+SAFB1/2 mutant settings were far more separated from wild type along PC1 than ERH mutants, but the two genotypes with SAFB1/2 loss separated from each other on the PC2. We then assessed the behavior of miRNAs previously reported to be sensitive to ERH and/or SAFB function from meta-analysis of data collected from different cell types (Kwon et al. 2020). Our data largely agree with published data, though we generally found larger effect sizes for dependence on SAFB1/2 (Supplemental Fig. S7). We also note that the triple ERH+SAFB1/2 mutant condition appeared to have milder changes for several miRNAs than SAFB1/2 mutants. While this nominally represents a rescue, given that ERH and SAFB1/2 have numerous functional roles beyond miRNA cluster assistance, this might also reflect a combination of indirect consequences in these mutant cells. For example, our data also recapitulate that a few miRNAs are upregulated in ERH and/or SAFB1/2 mutants (Supplemental Fig. S7; Kwon et al. 2020), which we tentatively assigned as indirect effects.

Figure 6.

Global effects of ERH and SAFB1/2 on endogenous miRNAs. (A) PCA of RPM-normalized read counts across the replicate small RNA libraries from four genotypes of HEK293T cells. (B) MA plots showing all tested (gray) and all clustered significantly downregulated (adjusted P-value [Padj] < 0.05, log₂ fold change [LFC] <−1; blue) and upregulated (Padj < 0.05, LFC > 1; red) miRNAs with average normalized counts (baseMean) ≥100. MiRNAs that were tested by Northern blotting in this study are annotated and highlighted in black, miRNAs that were filtered in the final data set (see the Materials and Methods) are plotted transparently, and miRNAs with baseMean <10 are omitted for clarity. Clustered miRNAs are enriched among the downregulated miRNAs, especially upon loss of SAFB factors. (C) Violin plots of miRNAs whose absolute value changed by greater than twofold (|LFC| > 1) in one of the mutants, highlighting the preferential downregulation of clustered miRNAs. (D) UpSet plots showing the overlaps of differentially expressed miRNAs among the genetic manipulations. Loss of SAFB1/2 has the most substantial impact, but many miRNAs are codependent on ERH and SAFB1/2. (E) Correlations of ERH and/or SAFB1/2 dependencies with the efficiency of in vitro cleavage by recombinant Microprocessor+SRSF3. We analyzed 487 miRNAs that had both in vitro cleavage data and >50 raw reads in at least one sample. The positive correlation between these data sets, especially with SAFB1/2 dependency, indicates that these Microprocessor cofactors facilitate suboptimal Microprocessor substrates. Pearson and Spearman correlation coefficients are denoted by ρ and s, respectively. (F) Comparison of steady-state and nascent (<6 h) miRNA expression changes in mutants of Microprocessor cofactors with sufficient nascent reads (440 loci). Positive correlation between these populations supports that these cofactors have direct impacts on miRNA biogenesis. For reference, miRNAs whose expression was validated by Northern blotting in this study are labeled in black.

Because our data are deeper than previous assessments of Microprocessor cofactor depletions (Fang and Bartel 2020; Hutter et al. 2020; Kwon et al. 2020), we were able to assign more differentially expressed miRNAs. MA plots (Fig. 6B) and volcano plots (Supplemental Fig. S8) summarize that loss of SAFB1/2 with or without simultaneous ERH depletion affected many more miRNAs than ERH depletion alone. We selected several classes of miRNA expression changes for validation. Interestingly, for the most part, we observed similar changes in the corresponding miRNA* strands, consistent with these reflecting altered processing (Supplemental Fig. S9). The endogenous miRNAs that we examined by Northern blotting included ones that were preferentially downregulated in SAFB1/2 mutants (miR-15a, miR-29c, and miR-221), ones that were dependent on both ERH and SAFB1/2 (miR-15b, miR-23b, miR-30b, and miR-191), one that was increased in mutants (miR-7), and a control that was relatively unchanged in mutants (miR-16). These experimental tests validated the changes measured from deep sequencing (Supplemental Fig. S9). While altered miRNAs spread in both directions, we note that clustered miRNAs were prominently enriched among the downregulated miRNAs. This is more clearly visualized in violin plots of the respective mutant genotypes, where we stratified miRNA accumulation by cluster status (Fig. 6C). Overall, these data are consistent with the notion that these Microprocessor cofactors play specialized roles to support miRNA cluster assistance.

It was earlier concluded that most suboptimal miRNAs are codependent on ERH and SAFB1/2 (Kwon et al. 2020). We used UpSet plots to visualize the genetic dependencies of miRNAs, and these highlight that most downregulated miRNAs are dependent on SAFB1/2 regardless of whether or not ERH was present and that additional miRNAs were SAFB1/2-alone (Fig. 6D; Supplemental Fig. S10). On the other hand, only a small minority of downregulated miRNAs were changed only in ERH or ERH+SAFB1/2 depletion conditions. Although a caveat is that these are strong loss-of-function but not null conditions, due to cell lethality of these factors in HEK293T (Fig. 1A), these data support the notion that SAFB1/2 play larger roles for cluster assistance than ERH. Moreover, they tentatively suggest that SAFB factors impact miRNA biogenesis downstream from ERH, in line with our other data (Figs. 1, 5).

Profiling from ERH and/or SAFB1/2 mutant conditions reveals suboptimal miRNAs

As noted, these mutant profiling experiments do not guarantee that miRNA expression is altered due to defective cluster assistance. For example, miR-1226-3p was reproducibly downregulated ∼20% in ERH and 50% in ERH+SAFB1/2 mutants (Supplemental Table S1). Because this is the mature strand of a mirtron, which uses splicing to bypass Microprocessor for nuclear biogenesis (Berezikov et al. 2007; Westholm and Lai 2011), this change should not be due to cluster assistance per se. Thus far, the only direct assay of suboptimality of nuclear miRNA processing is to show that the biogenesis of a given hairpin can be increased when in proximity to an optimal neighbor. This readout shows that even some “solo” miRNAs are suboptimal Microprocessor substrates (Shang and Lai 2023). However, such assays are not yet scalable.

To gain insight into whether dependence on Microprocessor cofactors reflects suboptimality as Microprocessor substrates, we took advantage of high-throughput in vitro Microprocessor cleavage data (Kim et al. 2021). Such cleavage data have a limitation in that the reaction conditions of recombinant proteins and synthetic RNAs are not expected to be the same as in cells (Shang et al. 2020). We observed a mildly positive correlation of ERH dependency with Microprocessor cleavage efficiency (Fig. 6E, left panel). However, we observed stronger correlations upon comparing in vitro cleavage efficiency with genetic dependencies on SAFB1/2 or ERH+SAFB1/2 (Fig. 6E, middle and right panels). This provides evidence that our profiling data helped to classify functionally suboptimal miRNAs that are enhanced by Microprocessor cofactors (Supplemental Table S2).

We summarize examples of ERH- and/or SAFB1/2-dependent miRNAs bearing suboptimal features in Supplemental Figure S11. We selected mir-92a for validation as a locus strongly dependent on SAFB1/2 but largely independent of ERH. Its behavior fits with our structure–function data (Fig. 2), because pri-mir-92a-1 exhibits a clearly suboptimal lower stem structure. We tested its accumulation across our mutant cell lines using Northern blotting and confirmed strong reduction of mature miR-92a in SAFB1/2-depleted but not ERH-depleted cells (Fig. 7). This provides further evidence of the separable functions of these Microprocessor cofactors. In fact, for reasons unknown, ERH+SAFB1/2-depleted cells accumulate higher levels of miR-92a than SAFB1/2-depleted cells. The basis of this potentially antagonistic effect remains to be clarified, but there might be further complexities in cluster biogenesis. For example, we have documented cases of Microprocessor competition (instead of collaboration) within clusters (Shang et al. 2020; Shang and Lai 2023).

Figure 7.

Impact of miRNA cluster assistance on Microprocessor homeostasis. (A) The DGCR8 transcript contains two nearby hairpins that are cleaved by Drosha. These are formally classified as miRNAs but inefficiently generate mature miRNAs, suggesting that their major function may be to regulate DGCR8 levels during Microprocessor homeostasis. mir-3618 resembles a canonical hairpin, but its pre-miRNA is reported to be nuclearly retained, while mir-1306 is a clearly suboptimal hairpin with poor lower stem pairing. (B) Expression of small RNAs and pre-miRNAs from ERH and/or SAFB1/2 mutant libraries. Mature miR-3618 and pre-mir-1306 were negligibly expressed (<0.25 RPM in all cases, usually less) and thus were not reliably quantified. (C) Northern blotting in wild-type and mutant settings for Microprocessor cofactors as well as in deletion backgrounds for mir-3618. These tests validate the specificity of pre-mir-3618 signals and confirm the absence of mature miR-3618 (see whole blots in Supplemental Fig. S13). These blots also show that miR-92a is selectively and substantially decreased in any condition with depletion of SAFB1/2; pri-mir-92a exhibits a suboptimal lower stem. (D) Western blotting of Microprocessor components and its factors. DGCR8 and Drosha proteins are increased upon depletion of ERH and SAFB1/2. Deletion of mir-3618 increases DGCR8 and Drosha proteins even more robustly, but loss of this DGCR8 hairpin blunts the effects of ERH and SAFB1/2 loss. (E) Model for complex feedback regulatory interactions that maintain Microprocessor (MP) homeostasis, including the impact of ERH/SAFB1/2-mediated cluster assistance on DGCR8 mRNA cleavage by Microprocessor.

Another major reason that miRNAs might change in steady-state profiling experiments is altered stability as opposed to biogenesis defects. To address this question, we analyzed miRNAs using the SLAM-seq labeling information aspect of our data. We extended a statistical modeling framework for small RNA SLAM-seq data (Herzog et al. 2017; Reichholf et al. 2019) to additionally infer the portion of unlabeled reads that were exposed to 4sU from the distribution of detected 4sU-induced T-to-C conversions, allowing us to distinguish between newly synthesized and preexisting miRNAs. We recovered a substantial population of newly synthesized miRNAs, which enabled broad assessment of their expression changes. As a quality control, we observed that miRNA* strands showed higher labeling fractions than their corresponding mature strands at the 6 h SLAM-seq time point (Supplemental Fig. S12), consistent with their overall faster turnover following maturation of Argonaute complexes into the mature single-stranded form (Reichholf et al. 2019).

Having verified the quality of our SLAM-seq data, we compared miRNA expression in the newly synthesized and steady-state pools. These analyses revealed that differential miRNA expression was strongly correlated between all three mutant genotypes (ERH, SAFB1/2, and ERH+SAFB1/2) compared with control HEK293T cells (Fig. 6F). Accordingly, these data confirm a dominant impact of depleting cluster assistance factors on miRNA biogenesis. Altogether, our data sets expand our knowledge of ERH- and/or SAFB1/2-dependent miRNAs, including suboptimal loci.

Impact of suboptimal miRNA biogenesis on Microprocessor homeostasis

In our final assays, we sought to characterize a regulatory output that is modulated by cluster assistance. Several clusters bearing suboptimal miRNAs are associated with phenotypically relevant biology (e.g., mir-144/451, mir-15a/16, and mir-17∼92a). In such contexts, ERH and SAFB1/2 levels might impact the repression of mRNA targets of suboptimal miRNAs. However, there are clear challenges to connect how changes in miRNA levels might be relayed to changes in their targets, given (1) the distributed nature of miRNA target networks and (2) the fact that ERH and SAFB1/2 have multiple roles outside of cluster assistance.

We therefore considered whether Drosha-mediated mRNA cleavage might be a favorable context to document in vivo regulatory impacts of cluster assistance. The first and most robust example of this is embedded within reciprocal cross-regulatory interactions between the two Microprocessor components. Microprocessor (via the ribonuclease Drosha) cleaves hairpins within the DGCR8 transcript to repress DGCR8 protein levels, while DGCR8 stabilizes Drosha protein within the Microprocessor complex. This homeostatic circuit operates in both vertebrates (Han et al. 2009; Triboulet et al. 2009) and intact Drosophila (Kadener et al. 2009; Smibert et al. 2011). Although the layout of two DGCR8 hairpins in mammals and flies is conserved, the miRNAs themselves are not, suggesting that cleavage of DGCR8 mRNA is more fundamental than function of the encoded miRNAs.

In vertebrate DGCR8, mir-3618 is located within the 5′ UTR, while mir-1306 resides in the N-terminal-encoding region (Fig. 7A). When first studied, these hairpins were confirmed as Drosha substrates but already noted to be less robustly processed than canonical pri-miRNA substrates (Han et al. 2009; Triboulet et al. 2009). In fact, mir-1306 was noted to be less amenable to Drosha cleavage than mir-3618 even though currently only mir-1306 (but not mir-3618) is included in curated MirGeneDB annotations (Clarke et al. 2025). The structure of mir-1306 is particularly suboptimal, with a poor lower stem structure, a modestly paired duplex stem, and a small terminal loop (Fig. 7A).

Although our profiling data were aimed at characterizing small RNAs, we noticed that they captured a minor subset of pre-miRNA reads (Supplemental Fig. S6E). Although mature miR-3618 reads were negligible, in line with previous reports that pre-mir-3618 is predominantly retained in the nucleus (Han et al. 2009), there were in fact substantial pre-mir-3618 reads in our sequencing data. These revealed a mild reduction of pre-mir-3618 in ERH mutants but a prominent greater than twofold reduction in SAFB1/2 and ERH+SAFB1/2 mutant cells (Fig. 7B). With mir-1306, its pre-miRNA reads were ∼100-fold lower than with mir-3618 and could not be assessed. However, we clearly observed that miR-1306 requires SAFB-mediated cluster assistance, because both its mature and passenger strand reads were reduced fivefold to 10-fold in SAFB1/2 and ERH+SAFB1/2 mutant cells (Fig. 7B). Although effects of ERH on miR-1306 were more subtle than with SAFB1/2, we note that miR-1306 was the most significantly reduced miRNA in prior analysis of ERH knockdown cells (Kwon et al. 2020); however, its fold changes were similarly modest in that study.

We could not detect mature miR-3618 or miR-1306 by Northern blotting, even upon sensitizing with ectopically expressed constructs (Supplemental Fig. S13). However, we detected a band of the expected size for the pre-mir-3618 hairpin (Han et al. 2009), along with a shorter band. To establish their provenance, we used CRISPR/Cas9 to delete pre-mir-3618 from WT and SAFB1/2 mutant cell lines (Supplemental Fig. S2). This eliminated the ∼58 nt band but did not affect the ∼42 nt band, validating the former as pre-mir-3618 and the latter as a cross-reacting species (Fig. 7C).

We then examined the effects on the accumulation of Microprocessor proteins. Our hypothesis was that enhanced cluster assistance would increase DGCR8 transcript cleavage, thus compromising accumulation of DGCR8 and Drosha proteins. Indeed, ERH and SAFB1/2 mutant conditions upregulated both Microprocessor proteins, though the effects were greater for DGCR8 (Fig. 7D), suggesting that it is the proximal target of cluster assistance. However, it is also possible that this reflects processing inhibition of a miRNA-like hairpin in the Drosha mRNA (Lee et al. 2017; Mechtler et al. 2017).

To more directly test whether these changes were due to cluster assistance per se, we used our mir-3618 deletion cells. Here, we hypothesize that loss of the optimal miRNA partner should (1) compromise repression of DGCR8 protein by Microprocessor and (2) blunt the regulatory effect of the remaining suboptimal mir-1306 hairpin. This is precisely what we observed: DGCR8 and Drosha proteins were misexpressed in all mir-3618 deletion genotypes, while the effects of ERH and SAFB1/2 were reduced in the absence of mir-3618 compared with cells that carry this hairpin (Fig. 7D). Overall, we conclude that one of the most overt utilizations of miRNA cluster assistance via Microprocessor cofactors is actually to mediate cross-regulatory interactions between Drosha and DGCR8 during Microprocessor homeostasis (Fig. 7E).

Discussion

Complex and separable roles for ERH and SAFB2 in miRNA cluster assistance

Since the discovery of cluster assistance for suboptimal miRNA biogenesis (Fang and Bartel 2020; Hutter et al. 2020; Shang et al. 2020), further mechanistic insights have been limited. A major reason for this is the complexity of this process, which involves Microprocessor (Drosha and DGCR8), at least two cofactors (ERH and SAFB2), at least two miRNA hairpins, and dynamic interactions among all of these. The proposed working models are all based on interactions of these cofactors with Microprocessor components, including ERH–DGCR8 (Kavanaugh et al. 2015; Fang and Bartel 2020; Kwon et al. 2020) and SAFB2–Drosha (Hutter et al. 2020) interactions, along with ERH–ERH dimerization (Arai et al. 2005; Wan et al. 2005; Kwon et al. 2020), SAFB2 dimerization (Hutter et al. 2020; Korn et al. 2023), and ERH–SAFB2 interaction (Drakouli et al. 2017). Accordingly, many plausible models of miRNA cluster assistance involve higher-order complexes of Microprocessor (Shang et al. 2023; Kim et al. 2025). However, the various models cannot be distinguished from the reported data.

While we can recapitulate some of these interactions, our data indicated that many proposed interactions are weak or not stably detected. More importantly, we found that several reported protein–protein interactions are not essential for cluster assistance. It remains possible that these are transient interactions and/or play supporting roles in the process. For example, although we showed that ERH-mediated dimerization of DGCR8 (Kwon et al. 2020) is not essential for cluster assistance, we also verified that dimerization of DGCR8 is indeed essential for canonical and suboptimal miRNA biogenesis. In fact, as three strategies were proposed for DGCR8 dimerization involving ERH (Kwon et al. 2020), heme binding (Garg et al. 2024), and the RHED domain (Faller et al. 2007), there are evidently fail-safe strategies to ensure this state. Nevertheless, our data showing that certain protein–protein interactions can be bypassed during effective cluster assistance help to distinguish correlative attributes from causal ones. Importantly, we were able to delineate a series of conditional requirements for Microprocessor cofactors and showed that these differ between ERH and SAFB2. Our data support the notion that ERH and SAFB2 play separable roles in a sequence of events that underlie Microprocessor transfer during miRNA cluster assistance, with ERH playing a role in Microprocessor transfer between hairpins, and SAFB2 being critical for stable recognition and/or processing of the suboptimal hairpin per se.

During review/revision of our study, related works on miRNA cluster assistance were published (Jang et al. 2025) or preprinted (Aschenwald et al. 2025). While it was previously reported that ERH mediates cluster assistance by binding the N terminus of DGCR8 (Kwon et al. 2020), our study and work from the Herzog laboratory (Aschenwald et al. 2025) indicate that loss of the ERH-binding region of DGCR8 is compatible with cluster assistance. In addition, work from the Kim laboratory (Jang et al. 2025) concluded that SAFB2 operates during transfer of Microprocessor between hairpins, while our study supports that SAFB2 is intrinsically required at the suboptimal miRNA hairpin for its processing (Fig. 5E). Tests from the Herzog laboratory (Aschenwald et al. 2025) are consistent with the latter interpretation.

In our small RNA sequencing data, we found that certain “solo” miRNAs are preferentially dependent on ERH and not on SAFB2 (Supplemental Fig. S11). This may be taken to reflect roles of ERH in miRNA biogenesis beyond cluster assistance, as proposed in concurrent studies (Aschenwald et al. 2025; Jang et al. 2025). Currently, our tests do not distinguish against the possibility of indirect consequences on miRNA upon silencing ERH. However, our tests confirmed that solo, suboptimal miRNAs do exist and that their maturation can be enhanced by ectopic Microprocessor (Shang and Lai 2023) or by recruitment via ERH (Fig. 3I,J). Of note, a recent study showed that SAFB can recognize an RNA hairpin in the Nfib transcript, thereby recruiting Drosha to cleave and downregulate Nfib in neural stem cells (Forcella et al. 2024). Thus, SAFB might also have roles in hairpin processing beyond miRNA cluster assistance. Finally, we also found that some miRNAs are elevated upon depletion of ERH and/or SAFB1/2, such as miR-7 (Supplemental Fig. S9C). Whether this is indirect or reflects additional complexity of Microprocessor cofactors on miRNA biogenesis deserves further study.

Co-option of multiple regulatory factors for regulation of nuclear miRNA biogenesis

Both ERH and SAFB1/2 are multifunctional factors that are involved in numerous other regulatory processes. Moreover, some of their core features are preserved in species that lack miRNAs. For example, ERH dimerizes in Schizosaccharomyces pombe, where it assembles a complex with the YTH protein Mmi1 (Hazra et al. 2020). Similarly, mammalian ERH dimerizes and assembles with DGCR8 (Kwon et al. 2020). It is therefore useful to provide some perspective on other processes that utilize these factors.

ERH is well conserved across eukaryotes and ranges from 100 to 113 amino acids (104 amino acids in humans), nearly qualifying as a “microprotein.” It was originally identified as an enhancer of the rudimentary wing defect caused by a defect in pyrimidine biosynthesis using Drosophila genetics (Wojcik et al. 1994) and was later found in mammals, plants, and fungi (Kozlowski 2023). ERH resides in both the cytoplasm and nucleus (Pogge von Strandmann et al. 2001) and can be selectively recruited to subcompartments such as nuclear speckles (Banko et al. 2013). Accordingly, ERH has been linked to a diverse array of regulatory processes including transcriptional regulation, splicing control, and mRNA decay (Kozlowski 2023). Strikingly, a recent study reported that ERH was one of the most highly interconnected factors among RBP-centered complexes, extending far beyond previously reported SAFB1/2 and DGCR8 (Street et al. 2024). Thus, ERH qualifies as a genuine regulatory hub.

Intriguingly, the Caenorhabditis elegans piRNA (21U RNA) biogenesis pathway, which bears little relationship to other metazoan piRNA pathways except for the fact that the resultant small RNAs assemble with a Piwi class effector protein, has co-opted an ERH homolog. In particular, ERH-2 homodimerizes as a constituent of two regulatory complexes with overlapping content (Cordeiro Rodrigues et al. 2019; Perez-Borrajero et al. 2021). While ERH-2 binds PID-3 in the common PETISCO complex, ERH-2 is the adaptor for two additional partners (PID-1 and TOST-1) that bind in a mutually exclusive fashion through the same interface. PID-1 confers activity in 21U biogenesis, while TOST-1 is important for embryonic viability that is independent of piRNAs. These findings highlight how ERH proteins have been independently co-opted for regulation of small RNA biogenesis and how these miniature proteins surprisingly have sufficient “real estate” to coordinate multiple distinct protein complexes.

SAFB1 (also known as SAFB) and SAFB2 are highly similar paralogs that are arranged head-to-head on human chromosome 19 (Renz and Fackelmayer 1996; Townson et al. 2003). Not only can both factors homodimerize, they are also members of a select group of proteins with capacity for independent recognition of DNA and RNA, accomplished via their SAP and RRM domains, respectively (Korn et al. 2023). As integral components of the nuclear matrix in vertebrate cells, both SAFB factors can bind scaffold/matrix attachment region (S/MAR) DNA. However, they also play significant roles in RNA regulation, including splicing and binding of nascent RNA (Townson et al. 2003; Sergeant et al. 2007; Rivers et al. 2015; Cherney et al. 2023). They are collectively involved in cell cycle regulation, apoptosis, differentiation, the stress response, regulation of immune genes (Norman et al. 2016), and transposon control (Hong et al. 2024; Ilık et al. 2024).

Overall, it seems sensible that ERH and SAFB1/2 could be co-opted into the miRNA pathway, because it is already clear that these versatile factors integrate into many other pathways. We showed that they have separable roles during cluster assistance, which may also be logical if we imagine that they were independently incorporated into this process. Thinking more broadly, we can speculate that even though the core of dedicated miRNA biogenesis factors was characterized long ago, there are still abundant opportunities for other proteins with other established functions to “moonlight” as part of the miRNA pathway.

Broader impacts and evolutionary implications of cluster assistance

Cluster assistance facilitates the processing not only of suboptimal miRNAs in mammals but also for miRNAs in Drosophila and some viruses (Haar et al. 2016; Truscott et al. 2016; Vilimova et al. 2021). Currently, only about a dozen suboptimal miRNAs whose biogenesis is enhanced by proximity to an optimal neighbor have been experimentally validated. As more than one-third of animal miRNAs reside in clusters, presumably more await characterization. Importantly, loss of ERH and SAFB1/2 can perturb miRNA expression independently of cluster assistance per se, and there may be additional strategies to enhance suboptimal miRNAs. Thus, direct manipulations of miRNA loci remain critical to document such regulation. A high-throughput strategy involving massive parallel screening would be critical to systematically reveal all the suboptimal miRNAs in different tissues and species.

The other intriguing open question is why cluster assistance is selected and kept for miRNA processing during evolution. One possibility is that young miRNA genes might be expected to be suboptimal, as de novo loci would not have been evolutionarily selected for optimal biogenesis features. Emergence within a cluster could facilitate the maturation of such a suboptimal miRNA (Mohammed et al. 2014; Shang et al. 2020) and permit evolutionary selection for beneficial targeting that could in turn promote enhanced biogenesis of the miRNA (Bartel and Chen 2004). This scenario may help explain young suboptimal miRNA hairpins that reside in miRNA clusters. However, other suboptimal miRNA genes, like mir-451 and mir-15a, are conserved across vertebrates yet maintain their suboptimal features. Why have these not transitioned into optimal miRNA hairpins? One generic explanation is that this allows greater control over miRNA biogenesis and thus gene regulation. However, a more specific rationale is that this enables hierarchical control of miRNA biogenesis. As we showed, suboptimal miRNAs are processed after their optimal partners (Shang et al. 2020; Shang and Lai 2023), and there might be settings in which tuning of cluster assistance mechanisms could enable selective expression of the optimal members of a miRNA operon. Finally, by virtue of its atypical Dicer-independent biogenesis strategy, mir-451 may be locked into utilizing cluster assistance. In fact, both mir-144 and mir-451 are subject to both independent and intertwined layers of biogenesis regulation (Yoda et al. 2013; Kretov et al. 2020; Shang et al. 2022), which may be relevant for their functions during erythropoiesis.

It is likely that other factors and strategies for enhancement of nuclear miRNA biogenesis exist. For example, our data indicate that ERH does not require direct interaction with DGCR8 to support cluster assistance via Microprocessor transfer. It is conceivable that this involves one or more other factors documented among the extensive ERH interaction network (Street et al. 2024). In addition, we previously showed that the solo miRNA mir-491 harbors a functionally suboptimal terminal loop that impedes Microprocessor recruitment (Shang and Lai 2023). While the biogenesis of pri-mir-491 can be strongly enhanced by a neighboring optimal miRNA, this strategy of cluster assistance is not available in its endogenous genomic location. We speculate that pri-mir-491 may recruit another unknown factor that facilitates recruitment and/or cleavage by Microprocessor. Indeed, there could be many miRNAs that utilize trans-acting factors whose identification and functional delineation await characterization (Treiber et al. 2017; Nussbacher and Yeo 2018).

Materials and methods

Plasmid constructs

Plasmids for expression of all the miRNAs or miRNA clusters used in this study were constructed by inserting amplified DNA fragments containing the miRNA precursors (200 bp to 1 kb) from genomic DNA of HEK293T cells between BglII and XhoI sites downstream from a CMV promoter. All the miRNA mutants were constructed using overlapping PCR method based on the wild-type miRNA constructs. The cDNA plasmids used in this study were constructed with HA, FLAG, and/or λN peptides on their N-terminal or C-terminal ends. The shRNA plasmids used in this study were constructed by directly inserting the annealed DNA oligos downstream from a U6 promoter (Shang et al. 2015). The CRISPR plasmids used in this study were constructed by directly inserting the annealed DNA oligos into the lentiCRISPRv2 vector with a GFP- or puromycin-resistant selection marker. All the details and oligonucleotide sequences used to clone these constructs are listed in Supplemental Table S3.

Cell culture

Wild-type and mutated HEK293T cells were grown in DMEM-high glucose media containing 10% FBS, 1% nonessential amino acids, 1% sodium pyruvate, penicillin/streptomycin, and 0.1% 2-mercaptoethanol. Mycoplasma contaminations were regularly tested for the cell lines.

RNAi-mediated knockdown

To deplete remaining ERH or SAFB1 factors, we applied RNAi-mediated depletion to the relevant mutant HEK293T cells. We transfected 2 µg of shRNA plasmids targeting the endogenous genes or control shRNA plasmids using Lipofectamine 2000 (Thermo Fisher) in 6 well plates. Thirty-six hours after transfection, cells were split and seeded into new culture plates for further transfections according to the specific experimental design.

Generation of knockout HEK293T cell lines using CRISPR/Cas9

Plasmids for gRNA expression were constructed by inserting annealed DNA oligonucleotides containing guide RNA sequences targeting ERH or SAFB1/2 genes into the BsmBI site of lentiCRISPRv2-GFP or lentiCRISPRv2-PuroR. Two micrograms of these plasmids was transfected into HEK293T cells in 6 well plates using Lipofectamine 2000. After 16 h, 2 µg/mL puromycin was added to cells transfected with lentiCRISPRv2-PuroR. After 48 h, the cells transfected with lentiCRISPRv2-GFP were collected for cell sorting to split the GFP-positive single cells into 96 well plates. After 2 weeks of culturing, the cells were used for genotyping and then Western blotting to identify the deletion or depletion colonies. Oligo sequences for sgRNA and genotyping are listed in Supplemental Table S3.

Northern blotting

Cotransfection of miRNA cluster plasmids (2 µg/well for 6 well plate or 1 µg/well for 12 well plate) with control mir-375 plasmid (200 ng/well for 6 well plate or 100 ng/well for 12 well plate) was performed in HEK293T cells using Lipofectamine 2000. Two days or 3 days after transfection, total RNA was prepared from the cells using TRIzol reagent (Invitrogen). Equal amounts of total RNAs (10–15 µg) were mixed with 2× RNA loading dye, denatured for 5 min at 95°C, and then fractionated by electrophoresis on a 20% urea polyacrylamide gel in 0.5× TBE buffer until the bromophenol blue dye ran out of the gel. The gel was transferred to a GeneScreen Plus nylon membrane (PerkinElmer) at 300 mA for 1.5 h, UV-cross-linked with 120,000 µJ of energy, and baked for 30 min at 80°C. Blots were hybridized with γ-³²P-labeled DNA probes against mature miRNA sequences in hybridization buffer (5× SSC, 7% SDS, 2× Denhardt's solution) overnight at 42°C. We washed the membrane with nonstringent wash solution (3× SSC, 5% SDS, 10× Denhardt's solution) followed by two washes with stringent wash solution (1× SSC, 1% SDS). Each wash step was conducted for 30 min at 42°C. Membranes were sealed in plastic wrap and exposed to a phosphorimager cassette for 1–3 days. To reprobe a blot, we washed the blot with 1% SDS for 30 min at 80°C prior to hybridization with additional probe(s). All the probe sequences are listed in Supplemental Table S3.

Coimmunoprecipitation assays

Wild-type or mutant HEK293T cells were grown in 6 well plates and transiently cotransfected with plasmids (1 µg for each) of FLAG- or HA-tagged cDNAs of relevant genes. After 48 h, the cells were washed with phosphate-buffered saline (PBS) and harvested in lysis buffer containing 50 mM Tris-HCl (pH 7.4), 150 mM NaCl, 1 mM EDTA, 1% Triton X-100, and protease inhibitor cocktail. After rotation for 20 min at 4°C, each lysate was clarified by centrifugation at 14,000g for 15 min at 4°C. A total of 500 µL of supernatant was mixed with 10 µL of anti-FLAG M2 agarose beads (Sigma) or anti-HA magnetic beads (Thermo Fisher) for 2 h at room temperature or overnight at 4°C. The beads were washed four times with Tris-buffered saline (TBS) and used for Western blotting analysis.

Western blotting

Input/IP products or HEK293T cell lysates were separated on 4%–20% Mini-Protean TGX precast protein gels (Bio-Rad) and then transferred to PVDF membranes. The blots were probed for 2 h at room temperature or overnight at 4°C with the relevant antibodies: rabbit polyclonal anti-SAFB1 (1:2000; Thermo Fisher A300-812A), rabbit polyclonal anti-SAFB2 (1:2000; Thermo Fisher A301-112A), rabbit polyclonal anti-ERH (1:2000; Thermo Fisher PA5-21388), rabbit monoclonal anti-DGCR8 (3F5; 1:3000; Thermo Fisher MA5-24860), rabbit monoclonal anti-Drosha (D28B1; 1:3000; Cell Signaling Technology 3364), or mouse monoclonal anti-β-tubulin (1:2000; Developmental Studies Hybridoma Bank), followed by secondary antibody conjugated to horseradish peroxidase (1:5000). For FLAG- or HA-tagged proteins, the blots were directly probed for 2 h at room temperature with HRP-conjugated mouse monoclonal anti-FLAG (1:5000; Sigma A8592) or HRP-conjugated mouse monoclonal anti-HA (1:5000; Thermo Fisher 26183-HRP). Signals were detected with Amersham ECL Prime Western blotting detection reagent.

Thiol (SH)-linked alkylation for the metabolic sequencing of RNA (SLAM-seq) small RNA library construction

The wild-type, ERH^+/−, SAFB1^+/−/SAFB2^−/−, and ERH^+/−/SAFB1^+/−/SAFB2^−/− HEK293T cells cultured in 6 well plates were first treated by shRNA transfections to further deplete endogenous ERH and SAFB1 as shown in Figure 1 for 48 h. Each genotype had two biological replicates. Next, cells from each well were divided, and half was replated for further culture while the other half was collected for validation. The replated cells were cultured for another 24 h and replaced with fresh medium containing 400 µM 4sU to label the nascent cellular RNAs in the dark for 6 h. Then, cells were collected by TRIzol for total RNA extraction, during which reducing agent (SLAM-seq kinetics kit anabolic kinetics module, Lexogen M06124-2-0100) was added to maintain the 4sU-treated samples under constant reducing conditions.

Following addition of 4sU, all subsequent steps were protected from light until noted. The total RNA samples were treated according to the above SLAM-seq kinetics kit protocol. In brief, 35 µL of the IAA/OS/NP master mix (composed of 5 µL of 100 mM freshly prepared iodoacetamide [IAA], 25 µL of organic solvent [OS], and 5 µL of sodium phosphate [NP]) was mixed with 15 µL of RNA (5 µg) for each sample. The reaction was incubated for 15 min at 50°C before addition of 1 µL of stopping reagent (SR) and then mixed to terminate the reaction. At this point, exposure to light was allowed. RNA was precipitated by adding 1 µL of carrier substance (CS), 5 µL of sodium acetate (NA), and 125 µL of 100% EtOH.

The purified RNAs (1 µg) from above were mixed with spike-ins (Qiagen 331535) followed by small RNA library construction using the Qiagen QIAseq miRNA library kit (331502) according to the protocol. In brief, 3′ and 5′ adapters were ligated to the RNAs, sequentially followed by reverse transcription using RT primers containing unique molecular indexes (UMIs), cDNA cleanup, library amplification, and library cleanup. This kit incorporates modified oligonucleotides to eliminate adapter dimers during the library construction and bead-based cleanups to eliminate unwanted background, bypassing the need for gel purification and size selection during library construction. The final libraries were sent for QC analysis and sequencing by Illumina PE100.

Data analysis for small RNA SLAM-seq

Sequenced reads were analyzed with a Nextflow (Di Tommaso et al. 2017) pipeline that orchestrated public bioinformatics tools and custom Python scripts based on the rnalib v0.0.4 (Popitsch and Ameres 2024), pysam v0.23.0, pandas v2.2.3, numpy v1.26.4, and scipy v1.15.2 packages. Briefly, we first trimmed sequencing adapters from the reads (first mate only) with UMI-tools v1.1.3 (Smith et al. 2017), matching for the QIAseq adapter sequence (AACTGTAGGCACCATCAAT) and extracting 12 nt UMIs downstream. We then removed and counted reads that globally aligned to the 52 QIAseq spike-in sequences using BioPython v1.85 (parameters: open_gap_score = −1 and defaults otherwise). Reads <15 nt were dropped as too short. Reads that aligned with a length-normalized alignment score of at least 0.9 were considered as spike-in reads and assigned to the spike-in sequence with the highest score. The remaining reads were then mapped to a transcriptome created from miRBase “miRNA_primary_transcript” sequences (padded upstream and downstream by 10 nt; overlapping sequences were merged and combined to single contigs) using NextGenMap v0.5.5 (Sedlazeck et al. 2013) in SLAM-seq aware mode (parameters: ‐‐slam-seq 2 -very-sensitive -n 3 ‐‐strata ‐‐kmer 8 ‐‐kmer-skip 0 -s 0.0 ‐‐min-residues 18 ‐‐gap-read-penalty 50 ‐‐gap-ref-penalty 50 ‐‐gap-extend-penalty 50). We then removed soft-clipped sequences from the aligned reads (stemming, e.g., from untemplated additions to the miRNAs) and combined those with all unmapped reads from the previous step.

These preprocessed reads were then mapped to the human genome (GRCh38) using bowtie v 1.3.1 (parameters: -v 3 -k 100 ‐‐best ‐‐strata) (Langmead et al. 2009), and all reads with at least one mismatch that did not overlap with miRBase-annotated miRNA loci were filtered as “off-target” reads. Finally, we continued processing using the original trimmed sequences (i.e., including tailed sequences) and mapped again to the transcriptome using NextGenMap (parameters as above). The motivation for this off-target filtering procedure was to remove sequences from non-miRNA loci to avoid false positive misalignments while retaining true positive miRNA reads with large edit distance (e.g., due to SLAM-seq-induced T-to-C conversions or due to tailing) to mature miRNA sequences.

We then deduplicated those alignments based on the extracted UMI sequences using UMI-tools dedup (parameters: ‐‐method unique), removed all reads with >10 mismatches as potential mapping artifacts, and counted miRNA sequences using a custom Python script. We assigned strand-matching reads to pre-miRNAs if they matched the miRBase-annotated “miRNA_primary_transcript” annotations (with a tolerance of 10 nt upstream or downstream) or if the whole read was enveloped by these annotations and covered at least 50% of those. We counted strand-matching reads as miRNA reads if their 5′ end matched the annotated miRNA 5′ end with a tolerance of 5 nt upstream/downstream, if their 3′ end matched the annotated 3′ end with a tolerance up to 5 nt upstream and 15 nt downstream, and if the read covered at least 50% of the annotated miRNA interval. Before this check, we clipped prefix and postfix (tailing) bases at the first mismatch positions beyond the miRNA annotations. Counts of multimapping reads were normalized by the number of optimal alignment positions (1/NH).

Our small RNA libraries were prepared with synthetic spike-in controls, but analysis indicated variable spike-in recovery across wild-type replicates (Supplemental Table S1, spikein_counts tab). Moreover, if normalizing to the apparently lower amount of spike-in reads in wild-type libraries, this would suggest globally decreased miRNA levels in mutants. However, Northern blotting tests confirmed that this was not the case and rather that there were selective upregulation and downregulation of miRNAs when normalizing by total RNA amounts (Supplemental Fig. S9). Consequently, spike-in normalization was not applied in downstream analyses. Instead, read counts were further normalized to sequencing depth (reads per million miRNA reads [RPM]). In the clipped reads, we also called strand-specific T-to-C conversions and respective convertible positions (filtering for per-base qualities <20 and randomly skipping calls with a probability of 0.0006 that we estimated from mismatches in the sequencing data to account for randomly spread sequencing errors).

To calculate fractions of T-to-C conversions (FTCs) per miRNA, we calculated the ratio of converted over convertible bases within the first 18 annotated nucleotides of a miRNA. To calculate effective 4sU (4sU_eff) concentrations per sample, we fitted a zero-inflated binomial model {ZIB ∼ [0 with probability ψ and B(n; p) with probability 1 − ψ], where B is the binomial distribution with parameters n corresponding to the number of convertible positions, p corresponding to the probability of T-to-C conversions (4sU_eff), and 1 − ψ corresponding to the fraction of new reads; i.e., reads from miRNAs created after addition of 4sU}. We fitted this model to the distribution of conversions per miRNA read with different numbers of convertible positions within the first annotated 18 nt and excluded the first position due to an observed underrepresentation of conversions in our alignments. We fitted 10 times to subsamples of 50,000 reads, calculated the mean of the predicted binomial probabilities, and then calculated the median of these values over a range of convertible position parameters (n in [4; 6] based on the data distributions) (data not shown). We then used the fitted p (i.e., 4sU_eff) values to estimate the fraction of new reads per miRNA: We fitted a ZIB model to the respective T-to-C conversion distributions per miRNA using the same procedure as above (subsample size 10,000) while fixing the binomial probability to the estimated 4sU_eff values and extracting 1 − ψ (i.e., frac_new).

To calculate counts of new reads, we multiplied these fractions by RPM-normalized read count values. For calculating differential expression of pre-miRNA and miRNA reads, we used the pyDESeq2 package v0.5.2 (Muzellec et al. 2023); miRNAs were considered significantly DE if their adjusted P-value was <0.05 and their absolute log₂ fold change was >1. The final set of considered miRNAs was determined by filtering for a minimum (raw) read count of 50 in any of the considered samples. Additionally, we removed 69 miRNAs with T-to-C convertible alignment columns within the first 18 nt of the miRNA annotation that had a coverage >10 and >75% converted bases as potential mapping artifacts. Our final data set contained read count data for 713 miRNAs and fractions of new reads for 441 miRNAs; all final counts are in Supplemental Table S1. For comparison with in vitro Microprocessor pri-miRNA cleavage (Fig. 6E), we downloaded Supplemental Table S6-2 from Kim et al. (2021) comprising efficiency scores of pri-miRNA cleavage by recombinant Microprocessor+SRSF3 and integrated the data with our data set (Supplemental Table S2). All data were analyzed and plotted in iPython notebooks; data QC was conducted using miRTrace v1.0.1 (Kang et al. 2018), Samtools v1.18 (Danecek et al. 2021), and FastQC v0.11.9.

Data availability

The raw SLAM-seq data were deposited at the European Nucleotide Archive under accession number PRJEB101478. In addition, the raw pipeline results and analysis notebook for creating the figures were assembled as a Zenodo repository available at https://www.doi.org/10.5281/zenodo.17464307.

Competing interest statement

The authors declare no competing interests.