Recombinant protein expression

The Drosophila Egl–BicD and human dynein complexes were expressed recombinantly from baculovirus in Sf9 insect cells (Thermo Fisher Scientific; derived from the ovarian tissue of Spodoptera frugiperda fall armyworm) using a polycistronic MultiBac system73, as described previously34. A codon-optimized Egl coding sequence (isoform B; NM_166623) with a C-terminal TEV–ZZ affinity tag was synthesized (Epoch Life Sciences) and cloned into the pACEBac1 acceptor plasmid and Cre-recombined with a pIDC donor plasmid containing a codon-optimized coding sequence for BicD (NM_165220; Epoch Life Sciences). The recombined plasmid was incorporated into the baculovirus genome by transformation of DH10EMBacY cells. This genome was purified and used with FuGENE HD (Promega) to transfect adherent Sf9 cells. After ~96 h at 27 °C, transfection or infection of the majority of Sf9 cells was confirmed by YFP detection. The baculovirus was then amplified by using the supernatant from the transfection to infect a 50-ml suspension culture of Sf9 cells for ~96 h at 27 °C. The supernatant of this culture was used to infect 500-ml suspension cultures for final protein expression. Cells from these cultures were pelleted, frozen in liquid nitrogen and stored at −80 °C until processed for protein purification. All Egl variants were derived from the pACEBac1 construct described above by site-directed mutagenesis (EglED1mut; Cys35Tyr) or by Gibson assembly (New England Biolabs) of PCR amplicons that excluded the targeted coding sequence (Egl∆ED2: deletion of residues 94–167, Egl∆ED3: deletion of residues 187–259, Egl∆XAD: deletion of residues 744–816) with the exception of EglExomut (Lys657Ala, Lys659Ala, Arg686Ala and Arg687Ala), which was commercially synthesized (Epoch Life Sciences) and cloned into pACEBac1. The instability of Egl–BicD lacking the ExoHD resulted in insufficient protein concentration for use. Deletion of ED1 was not pursued as it is expected to compromise Egl stability by eliminating BicD association34. The human dynein complex was expressed analogously to Egl–BicD, with sequences encoding dynein 1 heavy chain (NM_001376.4) bearing an N-terminal ZZ–TEV–SNAPf tag cloned into pACEBac1 and the remaining subunits (dynein intermediate chain 2 (DIC2: AF134477), dynein light intermediate chain 2 (DLIC2: NM_006141.2) and the light chains (Tctex: NM_006519.2, LC8: NM_003746.2, Robl: NM_014183.3)) cloned into pIDC74.

Protein purification

Egl–BicD complexes and dynein were produced as described previously34, with purification steps at 4 °C. Sf9 cells were routinely confirmed as free of Mycoplasma using the MycoALERT kit (Lonza). Cells expressing recombinant complexes were suspended in lysis buffer (50 mM HEPES pH 7.3, 500 mM NaCl (100 mM NaCl for dynein), 10% glycerol, 1 mM DTT, 0.1 mM Mg-ATP, 2 mM PMSF and 1× cOmplete EDTA-free protease inhibitor cocktail (Roche)) and lysed by passage in a Dounce homogenizer with a tight pestle. The lysate was clarified by ultracentrifugation in a Type 70 Ti rotor (Beckman-Coulter) at ~500,000g, applied to prewashed (in lysis buffer) IgG–Sepharose FF resin (Cytiva) in a gravity-flow Econo column (Bio-Rad) and incubated with gentle rolling for 3 h. Flowthrough was collected by gravity and the protein-bound resin was washed twice with five column volumes of lysis buffer and twice with five column volumes of TEV buffer (50 mM Tris pH 7.4, 150 mM potassium acetate, 2 mM magnesium acetate, 1 mM EGTA–KOH pH 7.5 and 10% glycerol). The resin was then transferred to a 15-ml conical tube in a 15-ml final volume of TEV buffer and incubated overnight at 4 °C with TEV protease to cleave the protein complexes from the beads and ZZ affinity tag. Liberated protein complexes were collected by gravity flow from a fresh Econo column and concentrated to ~500 µl (Egl–BicD) or ~300 µl (dynein) using Amicon Ultra-4 (100-kDa molecular weight cutoff (MWCO)) concentrator units (Merck) before application to Superose 6 Increase 10/300 (Egl–BicD; Cytiva) or TSKgel G4000SWxl with guard (dynein; TOSOH Bioscience) gel-filtration columns run in GF150 buffer (25 mM HEPES pH 7.3, 150 mM KCl, 1 mM MgCl2, 0.1 mM Mg-ATP, 5 mM DTT and 10% glycerol) on an AKTA Purifier fast protein liquid chromatography (FPLC) instrument (Cytiva). Fractions containing the protein complexes were pooled and concentrated as above to final concentrations of ~1.5 mg ml−1 (Egl–BicD and dynein for single-molecule motility assays) or 4–5 mg ml−1 (Egl–BicD for MST assays), as determined with Bradford reagent (Pierce).

Native dynactin was purified at 4 °C from pig brain extracts, as described previously74. Three pig brains were homogenized in lysis buffer (35 mM PIPES–KOH pH 7.2, 5 mM MgSO4, 1 mM EGTA–KOH pH 7.5, 0.5 mM EDTA pH 7.4, 1 mM DTT, 2 mM PMSF and 1× cOmplete EDTA-free protease inhibitor cocktail (Roche)) using short bursts in a blender. The lysate was clarified at low speed at 38,400g in a JLA 16.250 rotor (Beckman-Coulter) and then at higher speed by ultracentrifugation at ~160,000g. The clarified lysate was filtered through a 0.45-µm-syringe-tip filter (Fisher) and loaded onto an SP-Sepharose (Cytiva) cation-exchange column (~250 ml bed volume) equilibrated in SP-buffer A (35 mM PIPES–KOH pH 7.2, 5 mM MgSO4, 1 mM EGTA–KOH pH 7.5, 0.5 mM EDTA pH 7.4, 1 mM DTT and 0.1 mM Mg-ATP) and run on AKTA Pure FPLC instrumentation. Lysate was fractionated by washing with six column volumes of buffer A and eluting with a two-phase linear salt gradient with SP-buffer B (SP-buffer A with 1 M KCl), in which the first phase increased the KCl concentration to 250 mM over three column volumes and the second phase further increased the KCl concentration to 1 M over an additional one column volume. Dynactin-containing fractions were pooled and diluted twofold with Q-buffer A (35 mM PIPES–KOH pH 7.2, 5 mM MgSO4, 1 mM EGTA–KOH pH 7.5, 0.5 mM EDTA pH 7.4 and 1 mM DTT) and filtered through a 0.22-µm-syringe-tip filter (Fisher) before loading onto a MonoQ 16/10 column (Cytiva) equilibrated in Q-buffer A and running on AKTA Pure FPLC instrumentation. After loading, the MonoQ 16/10 column was washed with five column volumes of Q-Buffer A and the bound protein eluted in a three-phase linear salt gradient with Q-buffer B (Q-buffer A with 1 M KCl), in which the first phase increased the KCl concentration to 150 mM over one column volume, the second phase increased the KCl concentration to 350 mM over ten column volumes and the third phase increased the KCl concentration to 1 M over one column volume. Fractions containing dynactin were pooled and concentrated in an Amicon Ultra-4 (100-kDa MWCO) concentrator unit (Merck) to a final volume of 300–500 µl before application to a TSKgel G4000SWxl with guard (TOSOH Bioscience) gel-filtration column running in GF150 buffer on AKTA Purifier FPLC instrumentation. Fractions containing dynactin were pooled and concentrated as above to a final concentration of 1–2 mg ml−1, as determined with Bradford reagent. All purified proteins were dispensed into single-use aliquots, flash-frozen in liquid nitrogen and stored at −80 °C.

RNA templates, synthesis and purification

RNA constructs for cryo-EM and MST were synthesized commercially (Horizon Discovery; see Supplementary Table 3) with the exception of the 151-nt RNA containing the hairy localization element, which was transcribed in vitro from a hairy 3′ UTR template18 (GenBank X15905), as described below. The TLS and hSL1 constructs (as well as their mutants) included 5′ and 3′ flanks of 8 nt (TLS) or 10 nt (hSL1) of the native sequence from K10 or hairy 3′ UTRs. RNAs for MST included a 5′-Cy5 label attached to stem loops by a linker of two adenosine nucleotides (with the exception of the TLS and hSL1 whose native 5′ flanking sequence served as the linker). For single-molecule motility assays with the TLS and hSL1, Cy3-labeled versions of these transcripts were used in addition to Cy5-labeled constructs.

K10 3′ UTR constructs for single-molecule assays and embryo injection were transcribed from and numbered according to the K10 cDNA (GenBank AY060415). The template for the 1,490-nt 3′ UTR construct (positions 1608–3061 of the cDNA) contained the poly(A) signal from the cDNA and an additional 38 bp that included an SP6 promoter from the backbone that was not counted in the numbering. TLS mutations within the K10 3′ UTR were introduced using a construct with engineered HindIII and NheI sites on either side of the TLS that was previously shown not to affect localization of mRNA in embryos51. Synthetic DNA oligos (Merck) encoding the desired TLS mutations and complementarity to the HindIII and NheI overhangs were ligated to the digested plasmid containing the K10 3′ UTR, thereby replacing the encoded TLS with the mutant sequence. The scrambled TLS is a nonlocalizing control incapable of making a stem-loop structure51. The hairy 3′ UTR corresponds to positions 1185–1845 of GenBank X15905 and is part of a 730-nt transcript used for injection. Templates of a mutant version of the hairy 3′ UTR in which hSL1 lacks bulges were synthesized commercially (GenScript). The 573-nt fragments of the I-factor RNA coding sequence containing the wild-type ILS or a version lacking bulges were transcribed from commercially synthesized template (Integrated DNA Technologies) and correspond to positions 2932–3498 of GenBank M14954.2. The 839-nt bicoid 3′ UTR RNA containing wild-type bcdSLV or a version lacking bulges was transcribed from a commercially synthesized template (GenScript) and corresponds to positions 1702–2536 of GenBank NM_169159.4. Full-length 1,718-nt gurken transcripts containing the wild-type GLS or a version lacking bulges were transcribed from commercially synthesized templates (GenScript) and correspond to positions 1–1718 of GenBank NM_057220.3.

LacZ RNAs with individual TLS and KSE elements or a combination of both were transcribed from a fragment of the LacZ gene that was comparable in length to the K10 3′ UTR and predicted to code for an RNA with minimal secondary structure, as determined by the greatest average ss-count in mfold predictions75 (https://www.unafold.org/mfold/applications/rna-folding-form.php). At a position analogous to the location of the TLS in the K10 3′ UTR (673 bp), the TLS and KSE sequences were added to the design and commercially synthesized (GenScript). For the transcript containing both the TLS and KSE, the linker between the stem loops was the native sequence between hSL1 and hSL2 (ref. 18). This template was also used to transcribe the 145-nt TLS-KSE construct for single-molecule motility assays.

Linear DNA templates with a T7 or T3 polymerase promoter were prepared by PCR or by restriction digestion at the 3′ end of the desired transcript, followed by agarose gel purification. RNA was synthesized using MEGAscript T7 (full-length K10; Thermo Fisher) and MEGAshortscript T7 (hSL1-hSL2 and TLS-KSE; Thermo Fisher) in vitro transcription kits or the mMESSAGE mMACHINE T7 or T3 in vitro transcription kit (Thermo Fisher) to make 5′-capped transcripts for injection into Drosophila embryos. Fluorescence labeling of in vitro transcribed RNA was achieved by stochastic incorporation of Cy3-labeled or Cy5-labeled UTP included in synthesis reactions at a ratio in the reaction mix of 1:4 labeled to unlabeled UTP for single-molecule motility assays or 1:9 labeled to unlabeled UTP for injection assays. Following in vitro synthesis for 2–4 h at 37 °C, template DNA was digested with DNaseI and RNA was purified by phenol–chloroform–isoamyl alcohol extraction (25:24:1; Ambion) followed by successive passage through two Microspin G50 columns (Cytiva). RNA was then precipitated with ammonium acetate and ethanol. RNA concentration was determined using a NanoDrop One spectrophotometer (Thermo Fisher) and RNA integrity was confirmed by agarose gel electrophoresis before the RNA was dispensed into aliquots and stored at −80 °C.

Cryo-EM sample preparation

For assembling Egl–BicD complexes with the TLS, hSL1, ILS and GLS, 1.5 μM of the full-length Egl–BicD complex was incubated with 25 μM RNA for 45 min at 4 °C in GF150 buffer containing 0.00125% IGEPAL (MilliporeSigma). For the Egl–BicD–bcdSLV and Egl–BicD–hSL1-hSL2 complexes, a truncation of BicD (residues 322–782) was used that excludes CC1. For Egl–BicD–bcdSLV, 0.75 μM of Egl–BicD(322–782) was incubated with 25 μM of bcdSLV RNA. For the Egl–BicD–hSL1-hSL2 complex, 0.75 μM of Egl–BicD(322–782) was incubated with 2 μM of hSL1-hSL2 RNA. The sample was then centrifuged at 12,000g for 2 min. Then, 3.5 μl of the supernatant was applied to freshly glow-discharged Quantifoil R2/2 300-square-mesh gold grids (Quantifoil) in a Vitrobot IV (Thermo Fisher) at 95% humidity and 4 °C, incubated for 10 s and blotted for 2 s before being plunged into liquid ethane.

Cryo-EM data collection and image processing

Cryo-EM samples were imaged using a FEI Titan Krios (300 kV) equipped with a K3 detector and an energy filter with a 20-eV slit size (Gatan). Automated data collection was performed using Thermo Fisher EPU. All data collection statistics can be found in Tables 1 and 2.

Egl–BicD–TLS complex

A total of 8,568 videos were acquired at a magnification of ×81,000 (1.059 Å per pixel) using a 100-μm objective aperture, with 40 frames per video and a total fluence of ~48 e− per Å2. Global motion correction and dose-weighting were performed in RELION 4.0 (ref. 76) using MotionCor2 (ref. 77) with a B factor of 150 and 5 × 5 patches. Patch-based contrast transfer function (CTF) estimation and initial processing steps were conducted in CryoSPARC78. Particles were initially picked using an ellipse (180 × 100 Å) as a reference. Subsequent two-dimensional (2D) classification was used to identify protein-like densities, which were used as references for an additional round of reference-based particle picking (Supplementary Fig. 3). Approximately 5 million particles were extracted with a box size of 200 pixels and a pixel size of 2.12 Å. Then, 2D classification was performed to select ~1.6 million particles belonging to classes with protein-like features. Ab initio reconstruction generated four three-dimensional (3D) references, which were subsequently used for heterogeneous refinement. Two of these classes (class 1 and class 2; Supplementary Fig. 3) displayed density corresponding to a coiled-coil region, RNA and distinct domains of Egl. Particles from these classes were selected for further processing in RELION 4.0 and 5.0 (ref. 79). A third class (class 3) was observed that lacked density for ExoHD and XAD. In this class, the RNA and its bound ED1 and ED2 exhibited notable flexibility, which precluded high-resolution 3D refinement.

Approximately 1 million particles from classes 1 and 2 of the heterogeneous refinement were reextracted in RELION 4.0 with a box size of 280 pixels and a pixel size of 1.059 Å. This was followed by 3D refinement using class 2 as the reference. This choice was informed by the observation that the RNA and associated Egl domains on either side of the BicD coiled coil exhibited flexibility relative to each other. To improve particle alignment, one of the ExoHD–XAD modules was omitted from the 3D reference and the mask. Local particle motion was subsequently corrected using particle polishing, during which the particles were reextracted with a 360-pixel box at a pixel size of 1.059 Å. Another round of 3D refinement was performed, followed by CTF refinement to refine per-particle defocus and per-micrograph astigmatism and to estimate beam tilt and trefoil parameters. A final 3D refinement was then conducted, resulting in a 3.1-Å-resolution consensus structure. To sort for conformational and compositional heterogeneity, 3D classification without alignment was performed, focusing on either side of the BicD coiled coil. Classification with a mask around the ExoHD–XAD module included in previous 3D refinements enabled sorting of conformations within the stable part of the complex. Two major conformations (structures A and B; Supplementary Fig. 3) were identified, differing slightly in the position of the TLS RNA and its bound ED2, ExoHD and XAD. Structure A, at 3-Å resolution, was used for analyzing interactions among the different components of the complex and for identifying sites for mutagenesis as it was the best-resolved structure.

In addition to focused 3D classification and refinement, we performed 3D classification with a mask around the full complex that also incorporated the ExoHD–XAD module excluded in prior 3D refinements. This facilitated sorting of Egl–BicD–TLS structures with either one or two ExoHD–XAD modules bound, as well as the conformations within these particle populations. Approximately 50% of particles contained two ExoHD–XAD modules, while the other particles had only one. Among the structures with two ExoHD–XAD modules, distinct conformations (structures C, D and E; Supplementary Fig. 3) were observed. These conformations shared consistent RNA-binding interactions across the domains but differed in the relative orientation of the RNA and its associated domains with respect to the BicD coiled coil. Structure C, at 3.4-Å resolution, was the best-resolved structure with all domains engaged with the RNA and was, therefore, used to describe the architecture of the complex in Fig. 1.

Egl–BicD–hSL1 complex

One dataset with 12,420 videos was acquired with 1.06 Å per pixel, 50 frames per video and a total fluence of ~50 e− per Å2 and another with 16,453 videos was acquired with 0.91 Å per pixel, 56 frames per video and a total fluence of ~52 e− per Å2. For each dataset, global motion correction and dose-weighting were performed in RELION 4.0 using MotionCor2 with a B factor of 150 and 5 × 5 patches. Patch-based CTF estimation and initial processing steps were conducted in cryoSPARC. Particles were picked using an ellipse (180 × 100 Å) as a reference and approximately 9.1 million particles were extracted in total at a box size of 90 pixels with a binning factor of 4. Then, 2D classification was used to select approximately 4.9 million particles that belonged to 2D classes displaying protein-like densities (Supplementary Fig. 4). Ab initio reconstruction was used to generate four 3D references, which were subsequently used for two rounds of heterogeneous refinement. The class displaying density corresponding to a coiled-coil region, RNA and distinct domains of Egl (class 1; Supplementary Fig. 4) was selected after each round of heterogeneous refinement. The selected particles were reextracted at a box size of 320 pixels (1.06 Å per pixel) or 360 pixels (0.91 Å per pixel), followed by another round of ab initio reconstruction. A subset of 560,115 particles, showing well-defined densities for Egl–BicD (classes 1 and 2; Supplementary Fig. 4), was selected for further processing in RELION 4.0 and 5.0.

Particles from both the datasets were merged at this stage and a round of 3D refinement was performed. Local particle motion was subsequently corrected using particle polishing. Another round of 3D refinement was performed, followed by CTF refinement to refine per-particle defocus and per-micrograph astigmatism and to estimate beam tilt and anisotropic magnification parameters. A final 3D refinement was then conducted, resulting in a 3.3-Å-resolution consensus structure. To sort for conformational and compositional heterogeneity, 3D classification without alignment was performed, focusing on either side of the BicD coiled coil. Classification with a mask around the ExoHD–XAD module included in previous 3D refinements enabled sorting of conformations within the stable part of the complex. One major conformation (structure A; Supplementary Fig. 4) was identified for this part of the complex and was resolved at an overall resolution of 3.2 Å. Additionally, classification with a mask around the ExoHD–XAD module excluded in prior 3D refinements facilitated sorting of Egl–BicD–hSL1 structures with either one or two ExoHD–XAD modules bound, as well as the conformations within these particle populations. Approximately 50% of particles contained two ExoHD–XAD modules, while the other particles had only one. Among the structures with two ExoHD–XAD modules, distinct conformations (structures B and C; Supplementary Fig. 4) were resolved at an overall resolution of 3.9–4.1 Å. These conformations shared consistent RNA-binding interactions across the domains but differed in the relative orientation of the RNA and its associated domains with respect to the BicD coiled coil. Structure A was resolved at the highest resolution and was, therefore, used for analyzing interactions.

Egl–BicD–ILS complex

A total of 8,009 videos were acquired at a magnification of ×81,000 (1.059 Å per pixel) using a 100-μm objective aperture, with 40 frames per video and a total fluence of ~47 e− per Å2. Patch motion correction and patch-based CTF estimation were performed in cryoSPARC. Particles were picked using an ellipse (180 × 100 Å) as a reference and approximately 4.4 million particles were extracted at a box size of 200 pixels and a pixel size of 4.24 Å. Then, 2D classification was used to select approximately 2 million particles that belonged to 2D classes displaying protein-like densities (Supplementary Fig. 5). Ab initio reconstruction was used to generate three 3D references, which were subsequently used for heterogeneous refinement. The class displaying density corresponding to a coiled-coil region, RNA and distinct domains of Egl (class 1; Supplementary Fig. 5) was selected for another round of ab initio reconstruction followed by heterogeneous refinement. A subset of 360,597 particles that had well-defined densities for Egl–BicD (class 1; Supplementary Fig. 5) was selected for further processing in RELION 5.0. Local particle motion was subsequently corrected using particle polishing, during which the particles were reextracted with a 360-pixel box at a pixel size of 1.059 Å. Then, 3D refinement was performed using a mask that excluded one of the ExoHD–XAD modules, focusing on the most stable regions of the complex. This yielded a 3.4-Å-resolution structure of the Egl–BicD–ILS complex (structure A; Supplementary Fig. 5).

To sort for compositional heterogeneity of the ExoHD–XAD module, heterogeneous refinement was performed, classifying particles into two groups: one with a single ExoHD–XAD module and another with two. Approximately 57% of particles contained one bound ExoHD–XAD module, while ~43% contained two. The 3D refinement of the class with two ExoHD–XAD modules resulted in a 3.7-Å-resolution structure (structure B; Supplementary Fig. 5). However, local resolution distribution indicated that the ExoHD–XAD modules exhibited flexibility relative to each other, leading to decreased local resolution in these regions when both modules were included in the mask for refinement. Because of its higher quality, the 3.4-Å-resolution structure was used for model building and to elucidate the binding mechanism of Egl to the ILS.

Egl–BicD–bcdSLV complex

A total of 22,750 videos were acquired at a magnification of ×81,000 (1.059 Å per pixel) using a 100-μm objective aperture, with 50 frames per video and a total fluence of ~50 e− per Å2. Patch motion correction and patch-based CTF estimation were performed in cryoSPARC. Approximately 13 million particles (200-pixel box; 2.12 Å per pixel) were initially picked using an ellipse (180 × 100 Å) as a reference (Supplementary Fig. 6). In parallel, 3.3 million particles were also picked using Topaz80 with a model trained using a small subset of 12,000 particles selected from 2D classification. Subsequently, a round of heterogeneous refinement using 3D references similar to those used for the TLS-bound structure was performed. Classes showing defined features for Egl–BicD–RNA were selected for a round of 3D classification without alignment. The particles coming from template-based picking and Topaz picking were then merged. After removal of duplicate particles, approximately 880,000 particles remained. Local particle motion was subsequently corrected in RELION 5.0, during which the particles were reextracted with a 360-pixel box at a pixel size of 1.059 Å. Another round of 3D refinement was performed, resulting in a 3.3-Å-resolution consensus structure. To sort for conformational and compositional heterogeneity, 3D classification without alignment was performed, focusing on either side of the BicD coiled coil. Classification with a mask around the ExoHD–XAD module included in previous 3D refinements enabled sorting of conformations within the stable part of the complex. One major conformation (structure A; Supplementary Fig. 6) was identified for this part of the complex after 3D classification. The selected particles were then used for 3D refinement followed by CTF refinement (beam tilt and trefoil parameters) and finally another round of 3D refinement, which resulted in a 3.4-Å-resolution structure. Additionally, classification with a mask around the ExoHD–XAD module excluded in prior 3D refinements facilitated sorting of Egl–BicD–bcdSLV structures with either one or two ExoHD–XAD modules bound, as well as the conformations within these particle populations. Approximately 37% of the particles contained two ExoHD–XAD modules, while ~63% had only one. Among the structures with two ExoHD–XAD modules, distinct conformations (structures B and C; Supplementary Fig. 6) were sorted after an additional round of 3D classification with the mask around the whole molecule and resolved at an overall resolution of 4.2–4.4 Å. However, local resolution distribution indicated that the ExoHD–XAD modules exhibited flexibility relative to each other, leading to decreased local resolution in these regions when both modules were included in the mask for refinement. Structure A was resolved at the highest resolution and was, therefore, used for analyzing interactions.

Egl–BicD–GLS complex

A total of 12,592 videos were acquired at a magnification of ×81,000 (0.91 Å per pixel) using a 100-μm objective aperture, with 56 frames per video and a total fluence of ~53 e− per Å2. Global motion correction and dose-weighting were performed in RELION 5.0 using MotionCor2 with a B factor of 150 and 5 × 5 patches. Patch-based CTF estimation and initial processing steps were conducted in cryoSPARC. Approximately 6 million particles were initially picked using an ellipse (180 × 100 Å) as a reference (Supplementary Fig. 7). Subsequently, a round of heterogeneous refinement using 3D references similar to those used for the TLS-bound structure (Supplementary Fig. 3) was performed. A class showing defined features for Egl–BicD–RNA was selected for a round of 3D classification without alignment. The best-defined 3D class was selected and two more iterations of heterogeneous refinement and 3D classification were performed while selecting the 3D class showing defined features for Egl–BicD–RNA at each step. This resulted in a total of 41,216 particles, which were then used to train a picking model in Topaz. The picked particles (~2.3 million) were then sorted using three rounds of heterogeneous refinement and 3D classification as for particles picked using an elliptical reference, resulting in a set of 66,615 particles. Both sets of selected particles were merged and duplicated particles were removed. For this complex, while there were classes from 2D classification that showed two ExoHD–XAD modules engaged with the complex, the low number of overall particles hindered the classification of complexes with one or two ExoHD–XAD modules in 3D. Therefore, after initial particle sorting, the processing was focused on resolving the complex while having only one ExoHD–XAD module within the mask. Then, 3D refinement in RELION 5.0 was performed followed by particle polishing and CTF refinement (per-particle defocus and per-micrograph astigmatism), resulting in a 3.9-Å-resolution structure, which was used to analyze interactions.

Egl–BicD–hSL1- hSL2 complex

A total of 44,979 videos were acquired at a magnification of ×81,000 (1.059 Å per pixel) using a 100-μm objective aperture, with 50 frames per video and a total fluence of ~50 e− per Å2. Initial image processing steps including particle picking and sorting were performed as described for the Egl–BicD–bcdSLV structure (Supplementary Fig. 6). A total of 1.14 million particles were obtained after initial particle sorting (Supplementary Fig. 10). Local particle motion was subsequently corrected in RELION 5.0, during which the particles were reextracted with a 360-pixel box at a pixel size of 1.059 Å. Another round of 3D refinement was performed, resulting in a consensus structure. To sort for compositional heterogeneity, 3D classification without alignment was performed using a mask encompassing the full complex. We found that ~60% (675,190) of the particles had Egl–BicD bound to both hSL1 and hSL2, representing the Egl–BicD–hSL1-hSL2 complex. The remaining particles displayed both stem-loop-binding sites occupied by hSL1 stem loops, resembling the Egl–BicD–hSL1 complex.

To sort for conformational heterogeneity within the Egl–BicD–hSL1-hSL2 complex, another round of 3D classification without alignment was performed using a mask encompassing the full complex. This identified a major conformation, which was used for further 3D refinement, yielding a map at 3.4-Å resolution. This structure was used to analyze the interaction of hSL2 with Egl–BicD.

Although the linker between hSL1 and hSL2 in the Egl–BicD–hSL1-hSL2 structure appeared flexible, we detected additional low-resolution density connecting the two stem loops (Supplementary Fig. 10c) that was absent in the structures with only hSL1 bound. This observation shows that both stem loops in the Egl–BicD–hSL1-hSL2 structure originate from the same RNA molecule (Supplementary Fig. 10c).

We did not observe 3D classes where the ExoHD–XAD module was bound to the hSL2 stem loop; in contrast, this module was consistently bound to only hSL1. In cases where both stem-loop-binding sites were occupied by hSL1 either in this dataset or in the Egl–BicD–hSL1 complex dataset, we observed structural classes with two ExoHD–XAD modules bound, in addition to those with only one (Supplementary Figs. 4 and 10). These observations, together with those presented in Fig. 5, indicate that hSL2 is not compatible with ExoHD–XAD binding.

Model building and refinement

For modeling Egl, the model of D. melanogaster Egl isoform B (UniProt Q9W1K4) generated using AlphaFold2 (ref. 38) (AF-Q9W1K4-F1) was obtained from the AlphaFold Protein Structure Database81. The individual domains (ED1, ED2, ExoHD and XAD) were fitted into the cryo-EM density using UCSF ChimeraX82 guided by the side-chain densities and the shapes of the domains. Given the flexible linkers between domains in Egl, their assignment to specific Egl chains was guided by multiple structural criteria. ED1 and ED2 positioned across the BicD coiled coil were assigned to the same Egl molecule on the basis of cryo-EM density at low threshold connecting these domains and a linker length compatible with this arrangement. The ExoHD–XAD module was assigned to the same chain as it is adjacent to ED2. This was supported by an interaction between residues 386–394 and the ExoHD, which reduces the intervening sequence between ED2 and ExoHD to approximately 217 residues. Additional folded elements within this region likely shorten the effective linker length even further, making it more likely that the ExoHD adjacent to ED2 belongs to the same chain. Although less favorable, it is possible that the ExoHD–XAD module comes from a different Egl molecule to their neighboring ED2.

For modeling BicD, an AlphaFold2 prediction of D. melanogaster BicD (UniProt P16568) was used. The model was fitted into the cryo-EM map using UCSF ChimeraX guided by the side-chain densities and an AlphaFold2 prediction of ED1 (residues 1–82) bound to BicD residues 700–782.

For modeling the RNA stem loops, 3D structures predicted by AlphaFold3 server83 or those generated by RNAComposer server84 were used as starting models.

After placing individual protein domains or RNA stem loops into the cryo-EM map, the model went through iterative cycles of restrained flexible fitting using ISOLDE85, followed by user-guided refinement in ISOLDE or Coot86. Final model refinement and model validation were performed in PHENIX87. All refinement statistics can be found in Tables 1 and 2.

AlphaFold2 prediction

All structure predictions unless specified were performed using AlphaFold2 (ref. 38) (for single chains) or AlphaFold2-Multimer88 (for multiple chains) through a local installation of ColabFold89. Both AlphaFold2 and AlphaFold3 accurately predicted the interaction between ED1 and BicD. However, we were unable to obtain reliable predictions for the Egl–BicD complex bound to different RNA targets.

RNA secondary-structure prediction

RNA secondary structures were predicted using the RNAfold web server90 (http://rna.tbi.univie.ac.at/cgi-bin/RNAWebSuite/RNAfold.cgi). Some nucleotides predicted to be part of bulges or loops were found to be integrated into the dsRNA helix in the experimental structures, where they participated in stacking and noncanonical base-pair interactions (Supplementary Fig. 8b). Therefore, when generating the ∆b constructs, nucleotides were removed on the basis of experimental structures.

MST

The affinity of the Egl–BicD complex and its variants for RNA stem loops was measured using a NanoTemper Monolith instrument and MO.Control MST acquisition software (NanoTemper). Twofold serial dilutions of the Egl–BicD complex ranging from 1.59 × 10−10 M to 7.70 × 10−6 M (assuming 2:2 stoichiometry of Egl:BicD) were incubated with 1 nM Cy5-labeled RNA at room temperature for 15 min in GF150 buffer supplemented with 0.05% Tween-20. Higher concentrations of Egl–BicD could not be tested as they were prone to aggregation. Serial dilutions were loaded into standard capillaries (MO-K022) and irradiated with infrared light at room temperature for 10 s (medium MST power) at 20% excitation, with changes in fluorescence monitored by the pico-red detector. MST traces were analyzed in MO.Affinity Analysis with data fitting performed using 1.5 s on-time (Fhot = 0.5–1.5 s; Fcold = −0.5–0 s) and the Kd model, which was able to fit binding curves in the absence of saturated RNA binding given sufficient signal-to-noise ratio (minimum value of 5.0 to be considered interacting). Data were further analyzed in GraphPad Prism (version 10.4.0) to determine the s.d. of the fit values of Kd.

Injection of Drosophila embryos with fluorescent RNA

Wild-type embryos (w1118 strain; Bloomington Drosophila Stock Center, BL5905) were collected and injected with Cy3-labeled RNA as described previously18,46,51. Typically, up to 60 dechorionated embryos were mounted in Voltalef oil 10S for injection of syncytial blastoderms with a 250 ng µl−1 solution of RNA. The person performing the injections was blinded to the identity of the RNA being evaluated. Following the last injection, embryos were incubated at room temperature for 8 min (~13 min from injection of the first embryo) before fixation with formaldehyde-saturated heptane, removal of the vitelline membrane with fine syringe needles and mounting in Vectashield with DAPI (Vector Labs) for visualization of nuclei.

Imaging was conducted using a Zeiss LSM 710 or 780 confocal microscope (Supplementary Table 4) equipped with a ×40 (1.3 numerical aperture (NA)) oil-immersion objective. Laser intensity was adjusted to allow visualization of injected RNA without reaching saturation.

RNA localization efficiency was quantified by comparing in FIJI91 the apical and basal RNA intensity within uniform regions of interest (ROIs), whose size (7.391 µm2) was predefined as the mean area occupied by apically localized RNA at individual microtubule organizing centers (positioned just above nuclei) in ~100 embryos injected with wild-type K10 3′ UTR. For each image, four ROIs were placed in the apical region of the injection site, with each ROI centered on the brightest area of fluorescence intensity. Corresponding basal ROIs were translated vertically to a basal position just above the yolk. The mean fluorescence intensities of apical and basal ROIs were then averaged and background-corrected by subtracting the mean intensity of a basal region distant from the injection site before being expressed as a ratio of apical to basal intensity.

Single-molecule-resolution RNA motility assays

Total internal reflection fluorescence (TIRF)-based motility assays of reconstituted dynein transport complexes assembled with the indicated RNAs were performed as previously described34. Assembly mixes of 100 nM dynein, 500 nM Egl–BicD, 200 nM dynactin and 50 nM RNA (25 nM each of Cy3-labeled and Cy5-labeled samples) were incubated in a total volume of 5 µl of GF150 on ice for 1 h before imaging. Complexes were assembled with a tenfold molar deficit of total RNA relative to Egl–BicD to assess the sufficiency of single RNAs to activate motility. Assembly mixes were diluted 10–80-fold in motility buffer (30 mM HEPES pH 7.3, 50 mM KCl, 5 mM MgSO4, 1 mM EGTA pH 7.5, 1 mM DTT, 20 µM Taxol (Sigma), 0.5 mg ml−1 BSA and 1 mg ml−1 α-casein (Sigma)) with 1 mM Mg-ATP and an oxygen-scavenging system (1.25 μM glucose oxidase, 140 nM catalase, 71 mM 2-mercaptoethanol and 25 mM glucose) and applied to a ~10-µl flow chamber containing streptavidin-immobilized microtubules (labeled with HiLyte 488 and biotin porcine tubulin; Cytoskeleton) on a PEG-biotin-passivated cover slip. Motility of Cy3-labeled and Cy5-labeled RNA within these chambers was alternately recorded with an iXonEM + DU-897E electron-multiplying charge-coupled device camera (Andor) mounted on a Nikon TIRF system (Supplementary Table 4) with a Nikon APO TIRF ×100 (1.49 NA) oil objective using Micro-manager software92 at the maximum possible frame rate (~2 frames per s for each channel) and 100 ms of exposure for each channel. Samples were illuminated with a 150-mW Coherent Sapphire 488-nm laser, a 150-mW Coherent Sapphire 561-nm laser and a 100-mW Coherent CUBE 641-nm laser. Assemblies of dynein–dynactin–BicD–Egl complexes with Cy3-labeled and Cy5-labeled TLS-KSE constructs were imaged using a Nikon Ring-TIRF system (Supplementary Table 4) controlled by Micro-manager and equipped with the iLas 2 platform (GATACA Systems) and the same objective as above. These samples were illuminated for 200 ms with 488-nm, 561-nm and 647-nm lasers within a Cairn Multiline Kompact laser box and the motility in Cy3 and Cy5 channels was recorded simultaneously on independent Photometrics Prime 95B complementary metal–oxide–semiconductor cameras at the maximal frame rate (~4 frames per s). Colocalization of Cy3 and Cy5 RNA was analyzed manually in FIJI using kymographs derived from acquisitions described above.

RNA bend analysis

A custom Python program (https://github.com/carterlablmb/RNA_bend_analysis) was used to calculate the bend between the upper and lower helices of RNA localization signals. First, the Curves+ software93 was used to obtain a helical trajectory for each RNA stem loop. From these trajectories, the coordinates representing the upper and lower helical regions were separately selected and fitted to best-fit lines in 3D space using a singular value decomposition algorithm. Each best-fit line was defined by a centroid and a principal axis vector and the bend angle (in degrees) was determined by computing the dot product of the two principal axis vectors and then applying the inverse cosine.

Egl sequence conservation analysis and RNA alignments

Egl orthologs were identified by BLASTp