Structure of the Dicer-2–R2D2 heterodimer bound to a small RNA duplex

In flies, Argonaute2 (Ago2) and small interfering RNA (siRNA) form an RNA-induced silencing complex to repress viral transcripts1. The RNase III enzyme Dicer-2 associates with its partner protein R2D2 and cleaves long double-stranded RNAs to produce 21-nucleotide siRNA duplexes, which are then loaded into Ago2 in a defined orientation2–5. Here we report cryo-electron microscopy structures of the Dicer-2–R2D2 and Dicer-2–R2D2–siRNA complexes. R2D2 interacts with the helicase domain and the central linker of Dicer-2 to inhibit the promiscuous processing of microRNA precursors by Dicer-2. Notably, our structure represents the strand-selection state in the siRNA-loading process, and reveals that R2D2 asymmetrically recognizes the end of the siRNA duplex with the higher base-pairing stability, and the other end is exposed to the solvent and is accessible by Ago2. Our findings explain how R2D2 senses the thermodynamic asymmetry of the siRNA and facilitates the siRNA loading into Ago2 in a defined orientation, thereby determining which strand of the siRNA duplex is used by Ago2 as the guide strand for target silencing.

In flies, Argonaute2 (Ago2) and small interfering RNA (siRNA) form an RNA-induced silencing complex to repress viral transcripts 1 . The RNase III enzyme Dicer-2 associates with its partner protein R2D2 and cleaves long double-stranded RNAs to produce 21-nucleotide siRNA duplexes, which are then loaded into Ago2 in a defined orientation [2][3][4][5] . Here we report cryo-electron microscopy structures of the Dicer-2-R2D2 and Dicer-2-R2D2-siRNA complexes. R2D2 interacts with the helicase domain and the central linker of Dicer-2 to inhibit the promiscuous processing of microRNA precursors by Dicer-2. Notably, our structure represents the strand-selection state in the siRNA-loading process, and reveals that R2D2 asymmetrically recognizes the end of the siRNA duplex with the higher base-pairing stability, and the other end is exposed to the solvent and is accessible by Ago2. Our findings explain how R2D2 senses the thermodynamic asymmetry of the siRNA and facilitates the siRNA loading into Ago2 in a defined orientation, thereby determining which strand of the siRNA duplex is used by Ago2 as the guide strand for target silencing.
The Dicer-2-R2D2 heterodimer has critical roles in both siRNA production and siRNA loading onto Ago2. Dicer-2-R2D2 processively cleaves long dsRNA substrates in an ATP-dependent manner to produce 21-nucleotide (nt) siRNA duplexes 4,5 . Subsequently, Dicer-2-R2D2 re-associates with an siRNA duplex, which is then loaded into Ago2 with the aid of the Hsc70/Hsp90 chaperone machinery [17][18][19][20][21] . R2D2 contributes to determining the specificities in the siRNA production and siRNA loading. Dicer-2 processes pre-miRNAs inaccurately in vitro, but R2D2 inhibits the promiscuous pre-miRNA processing by Dicer-2 (ref. 5 ). Dicer-2-R2D2 efficiently binds highly paired siRNA duplexes, but not miRNA duplexes with central mismatches, thereby preventing the inappropriate loading of miRNA duplexes into Ago2 (ref. 22 ). Notably, Dicer-2-R2D2 binds an siRNA duplex in a fixed orientation: the more thermodynamically stable 5′ end of the siRNA duplex is located near R2D2, whereas the other 5′ end with the weaker base-pairing stability is positioned near Dicer-2 (ref. 3 ). In general, Ago2 uses the siRNA strand with the less thermodynamically stable 5′ end as the guide strand for target silencing 23,24 , whereas the other strand in the siRNA duplex is cleaved by Ago2 and discarded as the passenger strand 25 . Thus, the Dicer-2-R2D2 heterodimer senses the siRNA thermodynamic asymmetry and transfers the siRNA duplex into Ago2 in a defined orientation, thereby determining which strand of the siRNA duplex is used by Ago2 as the guide strand.
Article visible in the previous cryo-EM reconstructions at approximately 7 Å resolution 28 , and revealed that Dicer-2 comprises an amino-terminal helicase domain, a DUF283 domain, a platform-PAZ domain, two RNase III domains (RIIIa and RIIIb) and a carboxy-terminal dsRNA-binding domain (CRBD) (Fig. 1a-e and Supplementary Video 1). The helicase domain consists of the Hel1, Hel2 and Hel2i domains, and a pincer-like helix. The platform and PAZ domains are linked by a connector helix. The RIIIa and RIIIb domains form an intramolecular dimer to create the central RNase III active site. The RIIIa domain interacts with the connector helix and the DUF283-platform linker (Extended Data Fig. 2a), whereas the RIIIb domain interacts with the Hel1, DUF283 and CRBD domains (Extended Data Fig. 2b). Our high-resolution structures further revealed the presence of an α-helical domain inserted within the RIIIa domain (referred to as RIIIi) and a prominent linker region between the PAZ and RIIIa domains (referred to as the central linker) (Fig. 1a,d). The RIIIi domain interacts with the DUF283-platform linker and the platform domain (Extended Data Fig. 2c). Notably, the central linker is mostly ordered and extensively interacts with the eight domains (Hel1, Hel2i, Hel2, platform, PAZ, RIIIa, RIIIb and CRBD) of Dicer-2 and R2D2 ( Fig. 1d and Extended Data Fig. 2d-f). The central linker regions are highly conserved among the Dicer-2 orthologues, but not the miRNA-producing Dicers (human Dicer and Drosophila Dicer-1) (Extended Data Fig. 3a,b). R2D2 comprises two dsRNA-binding domains (RBD1 and RBD2) and a carboxy-terminal domain (CTD) (Fig. 1b,d). The three domains adopt dsRNA-binding domain folds with an αβββα topology.
The Dicer-2-R2D2-siRNA structure revealed two RNA duplex molecules: one bound to the helicase domain of Dicer-2 and the other bound to R2D2 (Fig. 1c-e, Extended Data Fig. 4a). The Dicer-2-bound RNA duplex was not well resolved in the density map (Extended Data Fig. 4b), suggesting that it does not bind stably to the helicase domain of Dicer-2. Since the density was ambiguous but fitted to the unstable end relative to the stable end of the RNA duplex, we modelled the nucleotides at the unstable end (nucleotides g1-5 and p15-21 in Fig. 1c) into the density. Nonetheless, the guide and passenger strands cannot be functionally defined at the dicing step 17 , so we do not discriminate between the two strands hereafter. By contrast, the R2D2-bound RNA duplex (except for nucleotides g21 and p21) was well resolved in the density map (Extended Data Fig. 4c), enabling us to unambiguously model the guide and passenger strands. These observations indicate that the RNA molecules bound to Dicer-2 and R2D2 represent a dsRNA substrate at the initial recognition state in the dicing process and an siRNA product at the strand-selection state in the loading process, respectively. Thus, we refer to the RNA molecules bound to Dicer-2 and R2D2 as dsRNA and siRNA, respectively.

Structural changes upon siRNA binding
A structural comparison of Dicer-2-R2D2 and Dicer-2-R2D2-siRNA revealed that, although their overall structures are similar, the central linker becomes ordered and interacts with Dicer-2 CRBD and R2D2 RBD2 upon siRNA binding (Extended Data Fig. 5a-c). R2D2 RBD1 in the Dicer-2-R2D2 structure was not resolved in the density map (Extended Data Figs. 1g and 5a), suggesting that RBD1 is highly mobile in the siRNA-unbound state. By contrast, RBD1 becomes ordered and interacts with RBD2 in the Dicer-2-R2D2-siRNA structure (Extended Data Figs. 1h and 5b-d), indicating a structural change in R2D2 upon siRNA binding.

dsRNA recognition by Dicer-2
The helicase domain of Dicer-2 adopts a C-shaped structure similar to that of RIG-I 34 , and contains a canonical ATP-binding site formed by P29-T35 (motif I), D139-C141 (motif II), and R494-R496 (motif VI) (Extended Data Fig. 6a-  shorter pincer-like helix (Extended Data Fig. 6b). In the Dicer-2-R2D2-siRNA structure, dsRNA is recognized by V67, G90, H147 and K177 in the Hel1 domain of Dicer-2 in a sequence-independent manner (Extended Data Fig. 6g), similar to the blunt-end dsRNA in the Dicer-2-dsRNA structure 28 (Extended Data Fig. 6h). These structural observations are consistent with previous studies indicating that the Dicer-2 helicase domain initially recognizes both long dsRNA substrates with a blunt end and a 2-nt 3′-overhanging end in the dicing process 28,35,36 . dsRNA cleavage by Dicer-2 The Dicer enzymes recognize the 5′-monophosphate and 3′-overhang of dsRNA substrates, using a basic pocket in the platform domain (5′-pocket) and a hydrophobic pocket in the PAZ domain (3′-pocket), respectively 27,30,31 . The present structure revealed that Dicer-2 has both 5′-and 3′-pockets similar to those of human Dicer and Arabidopsis DCL3 (Fig. 3a-c and Extended Data Fig. 7a-f). A previous mutational analysis indicated that H743 and R943 in the 5′-pocket are involved in siRNA production 37 . The RNase III domain of Dicer-2 is structurally similar to those of human Dicer 29 and Arabidopsis DCL3 31 , and contains the active sites formed by conserved acidic residues (E1213, D1217, D1368 and E1371 in RIIIa and E1472, D1476, D1614 and E1617 in RIIIb) (Fig. 3a-c). A structural comparison of Dicer-2-R2D2 with DCL3-dsRNA suggested that Dicer-2 recognizes the 5′-monophosphate of dsRNA substrates in the 5′-pocket and cleaves the dsRNAs 21 nt away from the 5′ end in the RIIIb active site (Extended Data Fig. 8a,b), consistent with a previous proposal 37 . The modelled dsRNA sterically clashes with the helicase and CRBD domains of Dicer-2 (Extended Data Fig. 8b), suggesting that these domains undergo structural rearrangements upon the binding of dsRNA substrates. Supporting this notion, in the Dicer-2 structure predicted by AlphaFold2 (ref. 38 ), the helicase domain is arranged similarly to that in the DCL1-dsRNA structure 30 and interacts with the DUF283 domain (Extended Data Fig. 8c,d). Furthermore, a comparison Guide-labelled siRNA duplex (g*) Passenger-labelled siRNA duplex (p*) U 5-Iodouracil * 32 P-radiolabelled phosphate Dicer-2 and R2D2 (WT or mutants) were incubated with 5′-radiolabelled siRNA (g* or p*) bearing 5-iodouracil at position 20. The reaction mixture was analysed by SDS-PAGE, and crosslinked proteins were detected using phosphorimaging. CTD Δα2, the R2D2 mutant in which F201 and H215 are connected by a GGGS linker. n = 3 independent experiments.

siRNA strand selection by R2D2
The present structure revealed that R2D2 fixes the siRNA duplex in a defined orientation (Fig. 4a,b and Extended Data Fig. 10a). The central region (nucleotides g5-15 and p5-15) and the stable end (nucleotides g16-20 and p1-4) of siRNA are recognized by R2D2. By contrast, the unstable end (nucleotides g1-4 and p16-20, containing a U-U pair) is not recognized by either the 3′-or 5′-pocket of Dicer-2, and instead is exposed to the solvent. The central region of siRNA is extensively recognized by RBD1 (Q11, R50, K52, R53, K56 and H57) and RBD2 (S124, P123, K145 and K147) of R2D2 through sugar-phosphate backbone interactions (Fig. 4c), consistent with a previous study showing that Dicer-2-R2D2 preferentially binds an siRNA duplex without central mismatches 22 . Notably, the 1-nt 3′-overhang at the siRNA stable end is anchored by the RBD2 and CTD of R2D2 ( Fig. 4d and Extended Data Fig. 10b). The 5′-phosphate group of the nucleotide p1 interacts with R101 and R150 of R2D2, consistent with a previous study showing that Dicer-2-R2D2 preferentially binds an siRNA duplex with a 5′-phosphate 3 . The ribose and nucleobase moieties of the nucleotide g20 stack with K98 and W205 of R2D2, respectively (Fig. 4d). Y204 in the second α-helix (α2) in CTD stacks with R101 in RBD2, stabilizing the RBD2-CTD interface.
To validate our structural findings, we performed photocrosslinking assays, using siRNA duplexes containing 5-iodouracil at position 20 of the guide strand (g*) or the passenger strand (p*) (Fig. 4e). The guide and passenger strands were crosslinked to R2D2 and Dicer-2, respectively ( Fig. 4f), as in a previous study 3 , consistent with our structural finding that the stable and unstable ends of the bound siRNA duplex are located in the vicinities of R2D2 (CTD) and Dicer-2 (PAZ), respectively. Whereas the K98A mutation did not affect the crosslinking with the siRNA duplexes, the W205A mutation and the CTD α2 deletion abolished the crosslinking of g* to R2D2, but not that of p* to Dicer-2 ( Fig. 4f), confirming that the siRNA stable end is located near W205. Notably, the CTD α2 deletion increased the crosslinking of g* to Dicer-2 (Fig. 4f). These results indicated that more siRNA duplexes bind to the CTD α2 deletion mutant in the opposite orientation, thereby highlighting the contribution of the CTD α2 to the asymmetrical siRNA binding. An siRNA duplex with a 1-nt 3′-overhang was similarly crosslinked to R2D2 (Extended Data Fig. 10c), consistent with the observation that U21 of the siRNA is disordered and not recognized by R2D2 in the present structure (Extended Data Fig. 10b). A blunt-end siRNA duplex was also crosslinked to R2D2 (Extended Data Fig. 10c). Consistently, the terminal base pair of a modelled blunt-end siRNA stacks with W205 and K208 of R2D2 (Extended Data Fig. 10d). These results indicated that R2D2 recognizes the double-helical conformation, rather than the 3′-overhang structure, of an siRNA duplex in the strand-selection process.
Together, our structural and functional data revealed that R2D2 prefers to bind the double-helical conformation at the end of an siRNA duplex in a sequence-independent manner. Consequently, the more thermodynamically stable end (with greater double-helical character) of the siRNA duplex is preferentially anchored by R2D2 in equilibrium, leading to the asymmetric recognition of the siRNA duplex by the Dicer-2-R2D2 heterodimer.

Discussion
We determined the high-resolution structure of the Dicer-2-R2D2 heterodimer bound to two RNA duplexes, which represent a dsRNA substrate in the pre-dicing, initial recognition state and an siRNA product in the pre-loading, strand-selection state. The structure provided mechanistic insights into dsRNA substrate recognition and siRNA thermodynamic asymmetry sensing by Dicer-2-R2D2. On the basis of the present structure, along with previous functional data, we propose a model of siRNA production and siRNA loading by the Dicer-2-R2D2 heterodimer ( Fig. 5 and Supplementary Video 2). The helicase domain of Dicer-2 recognizes a long dsRNA substrate and then undergoes a conformational change. The dsRNA substrate passes through the helicase domain, and the 5′ end of the dsRNA is anchored by the 5′-pocket in the platform-PAZ domain. The dsRNA substrate is cleaved in the RNase III active site, yielding 21-nt siRNA duplexes. The produced siRNA duplex is released from the active site, and then recaptured by R2D2. While the thermodynamically stable end of the siRNA duplex is recognized by R2D2, the 5′-phosphate of the siRNA guide strand is exposed to the solvent. Notably, Ago2 uses the MID domain to recognize the 5′-phosphate of the siRNA guide strand 39,40 , and the Hsc70-Hsp90 chaperone machinery facilitates the docking of Ago2 on the Dicer-2-R2D2-siRNA complex in a manner dependent on the recognition of the 5′-phosphate of the siRNA guide strand 21 . These observations suggest that Ago2 adopts an open conformation by the action of the Hsc70-Hsp90 chaperone machinery, and captures the 5′-phosphate of the guide strand in the siRNA duplex bound to Dicer-2-R2D2. In this way, the Dicer-2-R2D2 heterodimer senses the siRNA thermodynamic asymmetry and facilitates siRNA loading into Ago2 in a fixed orientation, thereby determining which strand of the siRNA duplex is used by Ago2 as the guide strand for target silencing. Future research should focus on the structural Article elucidation of the Dicer-2-R2D2-siRNA-Ago2 quaternary complex for a complete understanding of the RNA-induced silencing complex (RISC) assembly mechanism.

Online content
Any methods, additional references, Nature Research reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41586-022-04790-2.

Expression and purification of the Dicer-2-R2D2 heterodimer
Dicer-2 and R2D2 were co-expressed in Sf9 insect cells using the Bac-to-Bac system (Invitrogen). The gene encoding Dicer-2 (residues 1-1722) was cloned into a modified pFastBac vector (Invitrogen), in which the N-terminal 6×His tag was replaced with an 8×His-GFP tag. Dicer-2 was also cloned into the modified pFastBac vector, in which the N-terminal 6×His tag and the following TEV protease cleavage site were replaced with an 8×His tag, to improve the yield of the purified protein.
The gene encoding R2D2 (residues 1-311) was cloned into a modified pFastBac vector, in which the N-terminal 6×His tag was replaced with a 3×Flag tag. The sequences of the DNA oligonucleotides used for the vector construction are listed in Supplementary Table 2.

Cryo-EM sample preparation
The Dicer-2-R2D2 complex was concentrated to A 280 = 0.6, using a Vivaspin centrifugal filter device (100 kDa MW cut-off, Sartorius). The sample (3 μl) was applied to a freshly glow-discharged Cu 300 mesh R1/1 grid (Quantifoil), in a Vitrobot Mark IV (FEI) at 4 °C, with a waiting time of 30 s and a blotting time of 4 s under 100% humidity conditions. The Dicer-2-R2D2-siRNA complex was concentrated to A 280 = 0.9, using the Vivaspin centrifugal filter device. The sample (3 μl) was applied to a freshly glow-discharged Au 300 mesh R1/1 grid (Quantifoil), in a Vitrobot Mark IV at 4 °C, with a waiting time of 30 s and a blotting time of 4 s under 100% humidity conditions. The grids were plunge-frozen in liquid ethane cooled at liquid nitrogen temperature.

Cryo-EM data collection and processing
Cryo-EM data were collected using a Titan Krios G3i microscope (Thermo Fisher Scientific), running at 300 kV and equipped with a Gatan Quantum-LS Energy Filter (GIF) and a Gatan K3 Summit direct electron detector in the electron counting mode.
Micrographs for Dicer-2-R2D2 were recorded at a nominal magnification of ×105,000, corresponding to a calibrated pixel size of 0.83 Å at the electron exposure of 15.8 e − per pixel per s for 2.30 s, resulting in an accumulated exposure of 53 e − Å −2 . The data were automatically collected by the image shift method using the SerialEM software 41 , with a defocus range of −1.6 to −0.8 μm, and 2,745 movies were obtained and processed using RELION-3.1. From the 2,745 motion-corrected and dose-weighted micrographs, 1,688,210 particles were initially picked, and extracted at a pixel size of 3.66 Å. These particles were subjected to several rounds of 2D and 3D classifications. The selected 324,630 particles were re-extracted at a pixel size of 1.25 Å, and then subjected to 3D refinement, per-particle defocus refinement, beam-tilt refinement, Bayesian polishing 42 and 3D classification with the mask focusing on Dicer-2 CRBD and R2D2. The selected 144,979 particles were subjected to 3D refinement, and subsequent postprocessing of the map improved its global resolution to 3.3 Å, according to the Fourier shell correlation (FSC) = 0.143 criterion 43 . The local resolution was estimated by RELION-3.1.
Micrographs for Dicer-2-R2D2-siRNA were recorded at a nominal magnification of ×105,000, corresponding to a calibrated pixel size of 0.83 Å at the electron exposure of 15 e − per pixel per s for 2.30 s, resulting in an accumulated exposure of 48 e − Å −2 . The data were automatically collected by the image shift method using the SerialEM software, with a defocus range of −1.6 to −0.8 μm. In total, 3,663 movies were obtained, and the beam-induced motion correction, dose-weighting and CTF estimation were conducted similarly to those for Dicer-2-R2D2. From the 3,663 motion-corrected and dose-weighted micrographs, 2,181,396 particles were initially picked, and extracted at a pixel size of 4.15 Å. These particles were subjected to several rounds of 2D and 3D classifications. The selected 179,826 particles were then re-extracted at a pixel size of 0.99 Å, and subjected to 3D refinement, per-particle defocus refinement, beam-tilt refinement and Bayesian polishing. The particles were again subjected to 3D refinement, and subsequent postprocessing of the map improved its global resolution to 3.3 Å, according to the FSC = 0.143 criterion.

Model building and validation
The initial model of Dicer-2-R2D2 was built using Buccaneer 44 , and the model was then manually built using COOT 45 . The model of the Dicer-2-R2D2-siRNA complex was built based on the Dicer-2-R2D2 model. The density maps were improved with the DeepEMhancer program 46 . The models were refined using Servalcat Refmac5 (ref. 47 47 . In brief, the final models were 'shaken' by introducing random shifts to the atomic coordinates with a root mean squared deviation of 0.3 Å, and were refined against the first half map. The statistics of the 3D reconstruction and model refinement are summarized in Supplementary Table 1. The cryo-EM density maps were calculated with UCSF ChimeraX 51 , and the molecular graphics were prepared with CueMol (http://www.cuemol.org).

Reporting summary
Further information on research design is available in the Nature Research Reporting Summary linked to this paper.