Methods

Construct design
Wild type (WT) DRP1 isoform 2 sequence was purchased from DNASU (sequence ID HsCD00043627, UNIPROT identifier: O00429-3, also known as Dlp1a) and was cloned into pET16b plasmid (Novagen) between the Nde1 and BamH1 sites. The vector was kindly provided by the laboratory of Wesley I. Sundquist with a 10X-His tag followed by a PreScission protease site (Leu-Glu-Val-Leu-Phe-Gln-Gly-Pro). Wild-type MID49126-454 sequence was PCR amplified and cloned into pGEX6p1 vector having an N-terminal GST tag followed by a PreScission protease site. Site directed mutagenesis was performed on pET16b-DRP1 and pGEX6p1-MID49126-454 using the Gibson cloning method to introduce mutations51. All constructs were verified using Sanger sequencing.

Protein purification
Protein purification was performed as described52. Briefly, plasmids containing the WT DRP1 or MID49126-454 sequence were transformed in the BL21-DE3 (RIPL) strain of E. coli. The colonies were inoculated in LB culture medium and grown overnight. Secondary inoculations were done the next morning in ZY medium for auto-induction53,54. The cultures were grown to an OD600 of 0.8 at 37°C in baffled flasks and were shifted to 19°C to grow for another 12 hours. The cultures were spun down and the bacterial pellets were used for protein purification immediately or stored at −80°C.
Full length DRP1 WT and mutant variants were purified as described previously for DRP1 WT with modifications25. Briefly, the bacterial pellets were resuspended in buffer A (50 mM HEPES/NaOH (pH 7.5), 400 mM NaCl, 5 mM MgCl2, 40 mM imidazole, 1 mM DTT, 0.5 mg DNase (Roche) and protease inhibitors (10 mM pepstatin, 50 mM PMSF, 0.5 mM aprotinin and 2 mM leupeptin), followed by cell disruption with a probe sonicator. Lysates were cleared by centrifugation at 40,000×g in Beckman JA 25.50 rotors for 60 min at 4°C. The supernatant was filtered using a 0.45 μm filter and applied to Ni-NTA Agarose beads pre-equilibrated with buffer B (50 mM HEPES/NaOH (pH 7.5), 400 mM NaCl, 5 mM MgCl2, 40 mM imidazole, 1 mM DTT). Upon the application of the supernatant, the beads were washed with 20 column volumes each of buffer B and buffer C (50 mM HEPES/NaOH (pH 7.5), 800 mM NaCl, 5 mM MgCl2, 40 mM imidazole, 1 mM DTT, 1 mM ATP, 10 mM KCl) followed by buffer D (50 mM HEPES/NaOH (pH 7.5), 400 mM NaCl, 5 mM MgCl2, 80 mM imidazole, 1 mM DTT, 0.5% (w/v) CHAPS). A final pre-elution wash was done with 20 column volumes of buffer B. Bound DRP1 was eluted with buffer E (50 mM HEPES/NaOH (pH 7.5), 400 mM NaCl, 5 mM MgCl2, 300 mM imidazole, 1 mM DTT) and dialyzed overnight at 4°C against buffer B without imidazole in the presence of PreScission protease to cleave the N-terminal 10X-His tag. The protein was re-applied to a Ni-NTA column pre-equilibrated with dialysis buffer and was observed to bind the column without the 10X-His tag as well. Subsequently, the protein was eluted with buffer B containing 80 mM imidazole.
Pure protein was concentrated with a 30kDa molecular weight cutoff (MWCO) centrifugal concentration device (Millipore). In a final step, DRP1 was purified by size-exclusion chromatography (SEC) on a Superdex-200 column (GE) in buffer F containing 20 mM HEPES/NaOH (pH 7.5), 300 mM NaCl, 2.5 mM MgCl2 and 1 mM DTT. Fractions containing DRP1 were pooled, concentrated, flash-frozen as single use aliquots in liquid nitrogen and stored at −80°C. Exact masses for purified DRP1 proteins were validated by MALDI-TOF mass spectrometry.
MID49126-454 was purified as described with the following modifications37. pGEX6p1-MID49126-454 plasmid DNA (human, UNIPROT identifier: Isoform 1 Q96C03-1, also known as MIEF2) was transformed in BL21 (DE3) RIPL cells. The colonies were grown overnight in LB medium and secondary cultures were grown in ZY medium. Cells were grown to an OD600 of 0.8-1, collected by centrifugation and processed immediately or stored at −80°C as described above. The bacterial pellets were lysed as described above in MID-buffer A (50 mM Tris pH 8.0, 500 mM NaCl, 5% glycerol, 1 mM DTT and 0.1% (v/v) Triton X-100). The lysates were pre-cleared at 40,000×g and filtered using a 0.45 μm filter before applying to 3 ml glutathione sepharose beads (GE). After overnight binding to beads, the unbound protein was removed and the beads were washed using 20 column volumes each of MID-buffer A and MID-buffer B (50 mM Tris pH 8.0, 500 mM NaCl, 5% glycerol, 1 mM DTT). The protein was eluted with MID-buffer C (50 mM Tris pH 8.0, 500 mM NaCl, 5% glycerol, 1 mM DTT and 20 mM reduced glutathione). The eluate was cleaved overnight with PreScission protease while dialyzing against MID-buffer D (20 mM Tris pH 8.0, 100 mM NaCl, 5% glycerol, 1 mM DTT). Cleaved protein was further purified using ion-exchange chromatography using a Q sepharose (GE) column. The low salt buffer for ion-exchange was the same as MID-buffer D and the high salt buffer was MID-buffer E (20 mM Tris pH 8.0, 1 M NaCl, 5% glycerol, 1 mM DTT). The relevant MID49126-454 fractions were pooled, concentrated and further purified using an SEC column pre-equilibrated with MID-buffer F (20 mM Tris pH 8.0, 200 mM NaCl, 5% glycerol, 1 mM DTT). The fractions containing MID49126-454 were pooled, concentrated, flash-frozen in liquid nitrogen and stored as single use aliquots at −80°C.

Filament assembly, EM sample preparation, data acquisition and processing
To assemble DRP1-MID49126-454 filaments, the proteins were mixed to a final concentration of 2 μM each and kept for an hour at room temperature. The mixture was dialyzed against assembly buffer- 20 mM HEPES pH 7.5, 50 mM KCl, 3 mM MgCl2, 1 mM DTT and 200 μM GMPPCP with or without 0.2% octyl-glucopyranoside (Anatrace). The filaments were observed using negative stain TEM or cryoEM after vitrification. Under these conditions, the mutant DRP1S611D failed to coassemble with MID49, but upon further lowering the ionic strength to 25mM KCl, DRP1S611D displayed detectable but greatly reduced coassembly compared to WT protein. For vitrification, the sample was applied to Quantifoil holey carbon grids (R2/2) using a Vitrobot Mark III with 3.5 μl sample, 5 seconds blotting time and a 0 mm offset at 19°C and 100% humidity. Images were collected on an FEI T30 Polara operating at 300kV at a magnification of 31000X. Images were recorded on a Gatan K2 summit camera in super resolution mode for a final binned pixel size of 1.22 Å/pixel. The movies were dose fractionated, contained 30-40 frames, had a total exposure time of 6-8 seconds with 0.2 seconds per frame and a per-frame dose of 1.1 to 1.4 electrons/Å2. SerialEM was used to automate data collection55. The defocus range was 0.8-3 μm under focus. The data was motion corrected and dose-weighted using UCSF Motioncor256. CTF parameter estimation on the non-dose-weighted but motion-corrected stacks was done using CTFFIND4 and GCTF57,58.
Filaments were boxed using the program e2helixboxer.py from the EMAN2 suite59. Particle coordinates were used to extract discrete particles using RELION 1.3-1.460 and all further processing was done within the RELION suite. Multiple rounds of 2D classification identified the well-ordered segments. 3D autorefine was run using a customized Relion1.2 version with the IHRSR algorithm implemented61,62. The consensus helical structure was used to classify the particles without refining helical symmetry (using RELION 1.4), resulting in 2 major classes that differed slightly in rise and twist (Extended data figure 2c). Particles from each class were selected and independently refined again with helical RELION 1.2 and IHRSR. Analysis of these reconstructions revealed that each structure was comprised of three linear filaments that bundle together to form a structure that resembled a triangle in cross-section (Extended data figures 2–3). The vertices of the triangle are formed through asymmetric interactions between the G-domains in adjacent filaments. The triangular arrangement of the bundled helices is unlikely to correspond to a biologically meaningful architecture, and this structure cannot form if the MID49 receptor is embedded in the outer mitochondrial membrane.
To further improve signal-to-noise, each of the three filaments in each independent half-map was segmented, extracted, resampled on a common grid and summed using UCSF Chimera63–67. The respective symmetrized but unfiltered half maps from each class were again aligned to a common grid and summed according to the C2 symmetry axis of the DRP1 dimer. In a last step, relion_postprocess was used to add the resulting and fully symmetrized half maps (Extended data figure 2c). These half maps and the final summed map, with differential B-factor sharpening per region (Extended data figures 2c, 4–5), were used for atomic modeling using Rosetta as described below.
For the projection structure of the DRP1G362D rings, 2 μM protein was mixed 1:1 molar ratio with MID49126-454 and was allowed to sit at room temperature for an hour. The mixture was dialyzed against the assembly buffer (without detergent) overnight. The sample was collected after 12 hours and vitrified using ultra-thin 3 nm carbon support films (Ted Pella). For vitrification, a Mark III vitrobot was used with 3.5 μl sample, 0 mm offset, 100% humidity and 3.5 seconds blot time. The images were collected using an FEI TF20 microscope and SerialEM for automated data collection. The data were recorded with a Gatan K2 camera operating in super resolution mode to collect dose fractionated movie stacks with a final binned pixel size of 1.234 Å/pixel. 40 frames were collected per stack (0.2 seconds per frame and 1.42 electrons/Å2). The movie stacks were motion corrected and the parameters of the transfer function were estimated as described above. Approximately 2000 particles were picked manually for initial 2D classification in RELION 1.4 and these averages were used as templates for further particle picking by Gautomatch (http://www.mrc-lmb.cam.ac.uk/kzhang/). Final 2D averages of the entire rings versus quarter segments of the rings were computed using Relion1.4.

Liposome and Nanotube reactions
Liposomes were made as described before52. Briefly, 1,2-dioleoyl-sn-glycero-3-phospho-L-serine (DOPS) was purchased from Avanti Polar Lipids. DOPS dissolved in chloroform was dried under a steady stream of nitrogen and dried under vacuum for an hour. The dried lipid film was resuspended in n-hexane and dried again under nitrogen. The resulting lipid film was dried under vacuum for 4 hours and was finally resuspended in 20 mM HEPES pH 7.5 and 150 mM KCl. The same protocol was followed for making nanotubes, where the mixture contained 60% D-galactosyl-ß-1,1′ N-nervonoyl-D-erythro-sphingosine (Galactosyl Ceramide), 30% DOPS and 10% Ni2+-NTA DOGS (1,2-dioleoyl-sn-glycero-3-[(N-(5-amino-1-carboxypentyl)iminodiacetic acid)succinyl] Nickel Salt).
For assembly reactions of DRP1G362D over lipid, 0.5-2μM protein was incubated with liposomes or nanotubes for an hour and dialyzed against the assembly buffer without detergent.

Model building
The general procedure for atomic model interpretation and validation using Rosetta were performed as described68. To obtain an initial model for DRP1, the crystal structure of nucleotide-free DRP1 (PDB ID: 4BEJ)25 was used for the stalk region and DRP1 G domain-BSE structures bound to GMPPCP (PDB ID: 3W6O) were used for the G-domain and BSE regions. Density-guided model completion for DRP1 was carried out with RosettaCM69 using this hybridization of DRP1 crystal structures. A converged solution appeared from the low-energy ensemble of the complete models generated by RosettaCM. However, among the low-energy ensemble, residues 503-610 were found to be extremely flexible without cryoEM density constraints and therefore were omitted for further coordinate refinement. For MID49, the highly homologous mouse MID49 crystal structure (81.3% identity, PDB ID: 4WOY, Extended Data Fig. 5c-d)37 was used to generate a homology model using RosettaCM and used as the starting model.
To enable fragment-based, density-guided model refinement with missing residues (503-610, DRP1), Rosetta iterative local rebuilding tool was customized to disallow backbone rebuilding at breaks within a single chain. Multiple rounds of refinement were done for each component against one half map (training map), and the other half map (validation map) was used to monitor overfitting according to the detailed procedure described in Wang et al.68.
With the refined model of DRP1 and MID49, we further refined the model in the context of a full assembly that included 8 identical copies of each protein, Mg2+ and nucleotide which included all possible inter-domain molecular interactions in the filament (Extended data figure 5a-b). Pseudo-symmetry was used70 to enable and facilitate the energy evaluations of all neighboring interactions around the asymmetric unit (Green model, shown in Extended data figure 5a) for final model refinement of the full assembly. To this end, refinement was done against the training map. Finally, the half maps were used to determine a weight for the density map that did not introduce overfitting. Using the weight and with the symmetry imposed, the whole assembly of DRP1 and MID49 was refined in the full map, followed by B-factors refinement71. Finally, quantification of buried surface area and the number and nature of the bonds involved for each DRP1-MID49 interaction interface modeled by Rosetta were performed with the PISA server (http://www.ebi.ac.uk/pdbe/pisa/).
Visual evaluations of the model-to-map correspondence was carried out in UCSF Chimera using unfiltered and unsharpened maps, maps uniformly sharpened with a range of ad hoc B-factors, and maps processed with a model-based local sharpening and local low-pass filtering procedure to optimize contrast and the visibility of high-resolution features of the map72.
To build a molecular model for the closed 12-dimer DRP1 rings, we used the diameter, thickness, and angles revealed by the 2D cryoEM class averages of the DRP1G362D rings stabilized with GMPPCP. The atomic coordinates determined above using RosettaCM were used to build the ring in sections, first with repeating dimers of the interface-2 “X-shaped” stalk, then the BSE and finally the G-domains and the angles between these sections were iteratively adjusted until calculated projections of the molecular model corresponded with the features of the experimental projection densities. Both the top (Fig 5b-c) and the side view (Extended data Figure 10b) were used as constraints. The complete atomic model of ring was finally refined in Phenix73 to minimize clashes.