Structures of Respiratory Supercomplex I+III<sub>2</sub> Reveal Functional and Conformational Crosstalk

All processing steps were done using RELION 2.0 and 2.1 (Kimanius et al., 2016) unless otherwise stated. MotionCor2 (Zheng et al., 2017) was used for whole-image drift correction of each micrograph. Contrast transfer function (CTF) parameters of the corrected micrographs were estimated using Gctf and refined locally for each particle (Zhang, 2016). Automated particle picking in Relion resulted in ∼400 k particles after manual curating. The particles were extracted using 4002 pixel box and sorted by reference-free 2D classification followed by re-extraction at 5122 pixel box and initial 3D classification with a regularization parameter T of 4 and a 30 Å low-pass filtered initial model generated from the PDB coordinates of CI and CIII2 fit into the low resolution reconstruction of the ovine respirasome (Letts et al., 2016b). This initial processing resulted in ∼250 k particles of good quality, which separated into several classes containing SC I+III2, isolated CI or isolated CIII2 (Figure S4C). As the SEC trace of the amphipol-stabilized SC I+III2 was monodisperse (Figure 1), we concluded that the isolated CI and CIII2 particles were generated during the grid preparation and plunge-freezing process. The isolated CI and SC I+III2 particles were separated and further classified to remove poor particles. At this stage, several distinct classes of the SC and isolated CI were identified and refined, resulting in reconstructions with nominal resolutions of 4.2 Å for three states and 4.6 Å for a fourth state according to the gold-standard FSC criteria.

Nonetheless, it was clear that the relative orientation between the two arms of CI and between CI and CIII2 in SC I+III2 was variable, which resulted in lower resolution at the edges of the CI peripheral arm and CIII2 (Figure 7E; Figure S7) and likely limited the nominal resolution of the reconstructions. To improve the maps, focused refinements were independently performed around three regions of the SC: the CI peripheral arm, the CI membrane arm and CIII2 (Figure S6). For the CI peripheral arm and CI membrane arm refinements, isolated CI particles were also included. After initial 3D classification, all of the CI-containing particles (i.e., isolated CI and SC I+III2) were combined and aligned via the peripheral arm of CI (Figure S5A). An additional round of 3D classification focused on the peripheral arm was performed to remove any remaining poor-quality particles. Interestingly, a small class of particles was identified that lacked the NADH-binding N-module of the peripheral arm (Figure S5A). The N-module is added to CI in the final stages of assembly (Formosa et al., 2018, Guerrero-Castillo et al., 2017) and its association may be dynamically controlled in response to ROS (Guarás et al., 2016). A class of particles lacking the N-module was also observed in the bovine respirasome structure (Sousa et al., 2016), suggesting that a sub-population of CI lacking the N-module is a common feature in mammalian mitochondria.

When we used the best class of 178,121 particles for 3D-refinement around the CI peripheral arm, we achieved a reconstruction at 3.8 Å, which allowed for atomic modeling of the CI peripheral arm (Figures S5A and S5D; Table 1). This class of 178,121 particles also defined the set of high-quality CI-containing particles that we used for all further CI 3D classifications (see below). Next, we aligned these particles via the CI membrane arm and performed an additional 3D classification to define a best-matching set of CI membrane arm particles (Figure S5C). We further refined the best classes from this classification (174,334 particles) to generate a reconstruction at 3.9 Å resolution, again allowing for atomic modeling of the CI membrane arm (Figures S5C and S5D; Table 1). As these sets of particles contained both isolated CI and SC I+III2, we performed the focused refinement for CIII2 with the original set of 120,596 particles identified in the original SC 3D classification (Figure S5B). After an additional round of 3D classification, we identified a set of 102,314 particles that, once refined, generated a 3.9 Å reconstruction around CIII2 (Figures S5B and S5D). It was found that re-centering and re-boxing the particles around the CI peripheral arm and CIII2 using the program REP (D’Imprima et al., 2017) improved map quality. This was not the case for the membrane arm which sits very near to the center of mass of the SC.

Finally, to better separate particles based on the difference between the CI membrane and peripheral arms, we developed a focus-revert-classify strategy in which the particles were first aligned by the peripheral arm of CI and then separated by 3D-classification with a mask around the membrane arm of CI without angular or translational searches (Figure 5). By pre-aligning particles by the peripheral arm, the small differences in the positions of the membrane arms should be maximized, thus facilitating efficient separation during 3D-classification without rotational or translational searches (Figure 5A). The peripheral arm was chosen for initial alignment as it contains all of the FeS cluster co-factors of CI, which generate strong peaks in the cryoEM density map. This strategy using ten classes resulted in the separation of at least seven distinct states of CI: one closed state and six open states (Figure S6A). The closed state and open states 1-3 each had significantly fewer particles than the original four SC states identified, but refined to similar or improved nominal resolutions, indicating that a more homogeneous set of particles for each state had been identified by this classification strategy. The nominal resolution of the seven classes ranged from 4.1-6.5 Å according to the gold-standard FSC criteria. Two additional classes at worse than 7.0 Å resolution were also generated but not analyzed further due to their poorer quality. Since no rotational or translational searches were performed, one class consisted of particles that did not align well with any of the other classes or each other and hence averaged out resulting in very weak and noisy density.

All local resolution maps were generated using BlockRes in Relion2.0 (Kimanius et al., 2016).

Starting models for isolated ovine CI (Fiedorczuk et al., 2016) and bovine CIII2 (Iwata et al., 1998), corrected for the ovine sequence, were used. These models were split and fit into the highest-resolution focused refinement maps for separate atomic model building of the CI peripheral arm, CI membrane arm and CIII2. Improved density in many regions relative to the previous maps of isolated CI allowed for the building of improved atomic models in COOT (Emsley et al., 2010). Real-space refinement of the model was done in PHENIX (Adams et al., 2010) and group atomic displacement parameters (ADPs) were refined in reciprocal space. Due to the unique features of cryoEM density (Coulomb potential density) relative to the electron density generated from X-ray diffraction data, such as frequent missing side chains for acidic amino acids caused by radiation damage (Baker and Rubinstein, 2010), we implemented a novel refinement protocol. First, we investigated whether the default values for group ADP refinements established for X-ray diffraction may be too stringent given the uncertainty in side chain position caused by radiation damage. Due to the moderate resolution of the reconstructions, individual (per-atom) ADP refinement was not performed, but, rather, group ADPs (one value for main chain atoms and another for side chain atoms) were assigned.

ADP refinement in PHENIX is done using local sphere restrains for each atom ai with three adjustable parameters: the sphere radius, all atoms within this radius of atom ai will be used to constrain its ADP value (default value 5 Å); the distance power, this parameter determines the degree to which the distance between two atoms ai and aj within the sphere radius influences the restraint (X-ray refinement default value: 1.69); and the average power, which determines the degree to which the ADP is restrained by the average value of ADP between atom ai and aj (X-ray refinement default value: 1.03) (Afonine et al., 2005). As has been pointed out previously, the ADP can represent the convolution of two types of uncertainty: 1) the dispersion of the atom position in the molecules (which may be the result of temperature-dependent atomic vibrations or static disorder in a crystal lattice) and 2) the uncertainty of the experimenter’s knowledge about the atom position (Merritt, 2012). The second type of uncertainty can be taken care of by occupancy refinement, but that is problematic at current cryoEM data resolutions. Therefore, we choose to optimize ADP refinement specifically for cryoEM data. We found that the default values underestimated the ADPs for side chains with poor density (such as Asp and Glu) when the density for the main chain atoms was well defined. We therefore tested an array of APD refinement parameters and found that decreasing the sphere radius to 2.5 Å and increasing the distance power to 2.3 without changing the average power resulted in more reasonable, higher group ADPs for side chains with poorly defined cryoEM density while maintaining overall consistency throughout the structure.

Another complication arising from refinement of models with poor side chain density for specific residues is that, unless the ADP is calculated before coordinate refinement, the refinement will attempt to place the side chain into the nearest main chain density displacing the main chain atoms and distorting the local geometry. We therefore implemented a refinement protocol in which, after manual model building in COOT, a single macro cycle of ADP refinement was performed in PHENIX (using the new defaults above) before coordinate refinement was performed. This allowed the ADP for side chains with poor cryoEM density to rise and hence reduce their propensity to displace main chain atoms during global minimization. The single cycle of group ADP refinement was followed by three cycles of global minimization, followed by an additional cycle of group ADP refinement and finally three cycles of global minimization. This refinement protocol resulted in models with improved geometry, reasonable group ADP values and better correlation with the cryoEM maps. Additionally, the ADP values of the model agreed qualitatively more with the local resolution of the cryoEM reconstructions (Figure S7). The script to run the protocol is available from authors on request.

In previous work on isolated ovine CI, subunit NDUFA11 (B14.7), TMH4 of ND6, the C-terminal TMH16 of ND5 and the C-terminal half of the ND5 lateral helix were disordered (Fiedorczuk et al., 2016). Given that these regions were clearly resolved in the focused maps of the CI membrane arm (Figure S5C), we could now build their model at atomic resolution. Additionally, in the previous structure, subunit NDUFB8 (ASHI) was built mostly as poly-alanine (Fiedorczuk et al., 2016). In contrast, we now present it in full atomic detail here. We also made significant improvements to the models of subunits NDUFB4 (B15), NDUFB9 (B22) and NDUFB7 (B18). Besides these major improvements, we also achieved minor improvements throughout the structure due to the high quality of the focused refinements (Figure S5). Overall, the CI model presented here in the closed SC I+III2 is > 95% atomic and the remaining unbuilt residues are almost entirely found at disordered regions on the N- and C-termini of supernumerary subunits (Table S5). The overall structure of CIII2 is similar to what has been previously reported for isolated CIII2 and in the respirasome (Guo et al., 2017, Huang et al., 2005, Iwata et al., 1998, Letts et al., 2016b, Wu et al., 2016). One major difference in our structure is the lack of density for the single TM helix subunit UQCR11. UQCR11, which is located at the periphery of the Q-binding cavities on either side of the CIII2 dimer, is added late in CIII2 assembly (Smith et al., 2012), is not required for CIII2 activity and is easily lost by chromatographic de-lipidation of the complex (Huang et al., 2005, Hunte et al., 2000, Schägger et al., 1990). Although UQCR11 is considered an authentic subunit of CIII2, like many of the respiratory chain supernumerary subunits, its role is not well understood (Smith et al., 2012).

UQCR11 was identified in our mass spec analysis (Table S1), which suggests that it is either sub-stoichiometric or disordered.

Due to low resolution, the focus-revert-classify reconstructions at ≥ 7.0 Å were not used in detailed structural comparisons, and poly-alanine models were only rigid-body fit into the reconstructions at 5.7-6.5 Å. We refined atomic models only for the four sub-4.5 Å reconstructions: the closed state and open states 1-3 (Figure 5B; Figure S6A).

All reported p values are from two-tailed Student’s t test for independent measures of unequal variance. Resolution estimations of cryoEM density maps are based on the 0.143 Fourier Shell Correlation (FSC) criterion (Chen et al., 2013, Rosenthal and Henderson, 2003).

The accession numbers for the atomic coordinates reported in the paper are PDB: 6Q9D, 6Q9B, 6QE9, 6QBX, 6QC3, 6QC2, 6QC4, 6QA9, 6QC5, 6QC6, 6QC8, 6QC7, 6QC9, 6QCA and 6QCF.

The accession numbers for the EM density maps reported in this paper are EMDB: 4480, 4479, 4481, 4493, 4495, 4494, 4496, 4482, 4497, 4498, 4500, 4499, 4501, 4502, 4505, 4506 and 4507.