Data reporting

No statistical methods were used to predetermine sample size. The experiments were not randomized. The investigators were not blinded to allocation during experiments and outcome assessment.

Sample preparation

Mitochondria were isolated from ovine heart tissue according to procedure 3 of ref. 44 and stored at −80 °C. Before supercomplex extraction with digitonin, frozen mitochondria were thawed on ice and washed by resuspension to a final concentration of ~6 mg protein per ml by manual homogenization in milliQ (18 MΩ) water to which KCl was added to a final concentration of 150 mM. Next, the membranes were pelleted by centrifugation at 32,000g for 45 min, followed by a second wash with resuspension in buffer M (20 mM Tris-HCl pH 7.4, 50 mM NaCl, 1 mM EDTA, 10% v/v glycerol, 2 mM DTT and 0.005% PMSF) by manual homogenization to a concentration of ~4 mg protein per ml and centrifugation at 32,000g for 45 min. Finally, membranes were resuspended in buffer M at ~10 mg protein per ml and either refrozen for storage at −80 °C or used directly for preparation of supercomplexes.

Supercomplexes were isolated from the washed mitochondrial membranes by digitonin extraction followed by sucrose gradient ultracentrifugation, as described previously18,45 with slight modifications. Briefly, an aliquot of washed mitochondrial membranes, containing ~5 mg of total protein, was resuspended in 2 ml buffer MX (150 mM potassium acetate, 30 mM HEPES pH 7.7, 10% glycerol and 0.002% PMSF) by manual homogenization. To the mitochondrial suspension, 1 ml of a 3% digitonin solution was added giving a final detergent concentration of 1% and a detergent to protein ratio of 6:1 (w:w). The sample was then agitated by rotation at 4 °C for 30 min before centrifugation at ~16,000g for 10 min. The supernatant was then concentrated to 0.5 ml and applied to linear sucrose gradients (10–45% sucrose in 15 mM HEPES pH 7.7, 20 mM KCl) prepared on a BioComp Gradient Station. The gradients were spun at 130,000g for 21 h, then fractionated. The fractions were run on BN–PAGE using linear gradient gels (4–20% polyacrylamide) in order to visualize the protein content (Extended Data Fig. 1). Fractions containing respirasomes were run over a PD-10 desalting column equilibrated in 100 mM NaCl, 20 mM HEPES pH 7.7, 0.1% digitonin and allowed to stand for ~1 h at room temperature until a white precipitate formed from the excess digitonin. The precipitate was pelleted by centrifugation and the supernatant was concentrated to an absorbance of ~2.8 absorbance units (AU) at 480 nm, which corresponds to about ~2.0 mg ml−1.

Electron microscopy

Aliquots of 2.7 μl of the isolated supercomplexes were applied to Quantifoil Cu R0.6/1, 300 mesh holey carbon film grids, which were glow-discharged in air for 120 s at 25 mA. Using FEI Vitrobot MKIII, the grids were blotted for 32 s at 4 °C at 100% humidity, plunged into liquid ethane and stored in liquid nitrogen. The grids were loaded onto a FEI Titan Krios transmission electron microscope (MRC LMB, Cambridge, UK) operated at 300 kV. Images were collected using EPU software on a Falcon-II detector at a calibrated magnification of ×81,395 (pixel size of 1.72 Å) and a dose rate of 17.0 electrons per Å2 per second (Extended Data Fig. 1b). Each image was exposed for a total of 4 s and dose-fractionated into 69 movie frames. A defocus range of 1.0–5.0 μm was used. Three datasets were collected, two datasets of freshly prepared supercomplexes comprising 1,178 micrographs and one dataset of overnight incubated supercomplexes of 430 micrographs for a total of 1,608 micrographs.

The dataset for supercomplex particles prepared in the same fashion but with an overnight incubation at 4 °C before grid preparations was collected in order to determine if the ratio of the tight-to-loose respirasomes changes with time. In the original datasets, the distribution of particles was 51% tight, 26% loose and 22% CICIII 2 (of 29,659 total good particles after 3D classification). For the particles incubated overnight, the distribution was 26% tight, 49% loose and 25% CICIII 2 (of 12,474 total good particles). Both supercomplex datasets were combined for the calculation of the final structures.

Image processing

All processing steps were performed using RELION v.1.4 (ref. 46) unless otherwise stated. A subset of ~2,000 particles were picked manually from 4-s averaged images, extracted using 2962 pixel box and subjected to reference-free 2D classification. Representative 2D classes were then used as reference images for automatic particle picking of all micrographs. The automatically picked particles were manually screened to remove false positives and pick any additional particles that were missed, resulting in an initial dataset of 67,650. MOTIONCORR47 was used for whole-image drift correction of movie frames 1–32 of each micrograph (the remaining frames were not used subsequently). Contrast transfer function (CTF) parameters of the corrected micrographs were estimated using Gctf and refined locally for each particle48. The particles were extracted using 2962 pixel box and sorted by reference-free 2D classification. We selected 54,484 particles from good 2D classes for the 3D classification (Extended Data Fig. 1c, d), which was run for 15 iterations, using an angular sampling of 1.8°, a regularization parameter T of 8 and a 30 Å low-pass filtered initial model from a previous low-resolution structure of the respirasome19, with a soft mask around the model. 3D classification was then continued as above for 15 iterations with an angular sampling of 0.9° and finally for 20 iterations with an angular sampling of 0.5°. This resulted in three good classes (tight, loose and supercomplex I–III) and a class of bad particles with no clearly resolved features (Extended Data Fig. 1d).

A subset of 42,133 particles including the three good classes was selected for the first 3D auto-refinement. This subset of particles were re-extracted from the motion corrected micrographs with a 4962 pixel box (to allow for high-resolution CTF correction49), and all further refinement was performed using this box size. After initial auto-refinement, a particle-based beam-induced motion correction and radiation-damage weighing (particle polishing) was performed (Extended Data Fig. 2e, f)50. Auto-refinement of all the polished particles together resulted in a reconstruction at 6.1 Å overall resolution with an estimated angular accuracy of 0.8°. The particles were then split into their respective 3D classes (tight, 18,397 particles; loose, 13,910 particles; and CICIII 2 , 9,844 particles; Extended Data Fig. 1d) and auto-refinement was run for each class independently, resulting in final reconstructions at 5.8, 6.7 and 7.8 Å resolution with angular accuracies of 0.6°, 0.7° and 1.0°, respectively. All resolutions are based on the gold-standard (two halves of data refined independently) Fourier shell correlation (FSC) = 0.143 criterion51 (Extended Data Fig. 2d). FSC curves were calculated using soft masks around the protein and high-resolution noise substitution was used to correct for convolution effects of the masks on the FSC curves52. Prior to visualization, all maps were corrected for the modulation transfer function of the detector.

Local resolution analysis by Resmap53 revealed a range of resolution for each supercomplex reconstruction with the highest resolution in the core of complexes I and III (Extended Data Fig. 2). The tight respirasome architecture shows resolutions ranging from 5.0 Å in the core of CI to ~8–9 Å resolution on the distal end of CIV (Extended Data Fig. 2a). The loose respirasome architecture shows resolutions ranging from 6.0 Å in the core of CI to ~10–12 Å resolution for the distal end of CIV (Extended Data Fig. 2b). Supercomplex CICIII 2 shows 7.0 Å resolution at the core of CI ranging to ~15 Å at the peripheries of the CI and CIII matrix domains (Extended Data Fig. 2c). In Fig. 1 and Extended Data Figs 3, 4, 5, the maps have been carved in order to remove the detergent micelle and give a clear view of the transmembrane helices. Due to weaker density for CIV in Fig. 1a and Extended Data Fig. 3, it was contoured at a lower level and in Fig. 1b and Extended Data Fig. 4 it was filtered to 8 Å, whereas the maps for CI and CIII are filtered at 6.7 Å.

Model building

For the CIII and CIV models, available high-resolution structures of the bovine enzymes (PDB accession codes 1BGY (ref. 54) 1NTM (ref. 43) and 1V54 (ref. 39)) were used as starting models. Mutations to the ovine sequences were made in COOT55 manually, as only few changes were needed. Sequences for ovine COX7c and COX8b were not available in the online databases and hence the bovine sequences were used. The amino acid residues in these models were truncated at the beta-carbon using CHAINSAW56 in the CCP4 program suite57 to generate a ‘poly-alanine’ model and fit into the cryo-EM density map for the tight respirasomes as a rigid body.

For CI, the low-resolution poly-alanine model of the bovine enzyme was used as a starting model (PDB accession code 4UQ8 (ref. 23)). This model was fit into the tight respirasome map and manual building was performed in COOT55 using Ramachandran and secondary-structure restraints. Mammalian CI consists of 14 conserved core subunits present throughout the species, including bacteria, and 31 mitochondria-specific supernumerary subunits (29 unique supernumerary subunits and two copies of the acyl carrier protein SDAP). As noted before23, core subunits retain the fold and architecture originally characterized in bacteria29,58,59. Building was guided largely by secondary structure predictions and homology structure predictions generated using the programs PSIPRED60,61 and Phyre2 (ref. 62). We improved the completeness of the poly-alanine model for all currently assigned subunits (with the exception of the B8 and SDAP subunits whose structures are known from homologues63,64) and assigned three additional subunits (Extended Data Fig. 7).

B18 and PDSW are both found on the intermembrane space side of CI and both contain strong secondary structure predictions for helix-turn-helix motifs with B18 containing a double CX 9 C CHCH domain65. Two regions of density were previously identified as probable candidates for either of these subunits, but a definitive assignment between the two was not made66. Based on secondary structure prediction, which differentiates the subunits by their helical structure at the C terminus, we assigned B18 to the density near the interface of CI and IV that extends a long (~30 amino acid residues) helix into the interface of the three complexes (Fig. 2 and Extended Data Fig. 7). This means that the helical density found underneath ND4 and ND5 corresponds to PDSW, which is predicted to have an additional shorter C-terminal α-helix (15–20 amino acid residues) following its predicted helix-turn-helix motif. Density for an α-helix (which we suggest belongs to the PDSW subunit) can be seen near the end of the PDSW helix-turn-helix motif that protrudes from CI towards CIII in the supercomplex structures.

Additionally, we assigned and built subunit B17.2. B17.2 contains strong secondary structure prediction on its N terminus for two short helices followed by a 3–4 strand β-sheet, then by a long coil (~70 amino acid residues) with no predicted secondary structure. Starting from a model generated by homology structure prediction using Phyre2 (ref. 62), we were able to fit B17.2 into density adjacent to the 49-kDa, TYKY and PSST subunits of the CI Q-modules. After manual adjustment of the secondary structure elements from the predicted homology model, density could be seen extending away from the C terminus, which clearly belonged to the B17.2 C-terminal coil. These C-terminal residues snake along the surface of the peripheral arm and extend towards the N-module. This extended coil connecting the N- and Q-modules of the hydrophilic arm speaks to the role of B17.2 and its homologue B17.2L in CI assembly67.

In subsequent work with isolated ovine CI, we assigned all remaining CI subunits32 and, where relevant, these assignments (B15, B12 and AGGG) are used here.