Complex I was isolated from mitochondrial membranes prepared from whole adult Drosophila by detergent extraction from the membrane followed by anion-exchange and size-exclusion chromatography, according to a small-scale protocol developed previously for mammalian complex I (Agip et al., 2018). The complex eluted from the size-exclusion column in a homogeneous peak consistent with the expected ~1 MDa mass of the monomeric complex (Figure 1—figure supplement 1). The highest concentration peak fraction (3.4 mg mL^–1), which exhibited an NADH oxidoreductase activity comparable to mammalian complex I of 7.3 ± 0.3 µmol min^–1 mg^–1 (ca. 120 NADH s^–1), was collected and frozen onto thiol-modified gold cryo-EM grids (Blaza et al., 2018; Meyerson et al., 2015; Russo and Passmore, 2014). The grids were imaged using a 300 KeV Titan Krios microscope equipped with a Gatan K2 camera and GIF Quantum energy filter (Table 1), and 63,471 particles images were selected and processed using RELION (Zivanov et al., 2020; Zivanov et al., 2018) into three major classes (Figure 1—figure supplement 2). The highest resolution map reached an estimated global resolution of 3.3 Å with consistent local resolution, and the two smaller subclasses reached estimated global resolutions of 3.7 and 4.0 Å (Figure 1—figure supplement 3). Example densities are shown in Figure 1—figure supplements 4–6.

Figure 1 shows the overall structure of Drosophila complex I, which consists of 43 subunits: 14 core subunits (Figure 1a) and 29 supernumerary subunits (Figure 1b). The 14 core subunits comprise the canonical heart of the enzyme that is conserved throughout all species of complex I, with the core subunits of the Drosophila and mammalian (bovine, PDB ID: 7QSK; Chung et al., 2022b) enzymes exhibiting an overall root mean square deviation (RMSD) of 1.065 Å. The 29 supernumerary subunits all correspond to supernumerary subunits found in mammalian complex I, confirming the close relationship between them. However, two supernumerary subunits present in mammalian complex I are absent from the Drosophila complex (Figure 1c): subunits NDUFC1 and NDUFA2 (to aid comparisons to the mammalian enzyme, we use the human nomenclature throughout; see, e.g., Rhooms et al., 2020 for a list of the corresponding gene names in Drosophila). NDUFC1 is a short, single transmembrane helix (TMH)-containing subunit in the membrane domain that is peripherally associated with the mammalian complex through its interaction with subunit NDUFC2, and subunit NDUFA2 binds to subunit NDUFS1 at the top of the hydrophilic domain in the mammalian complex. The absence of NDUFC1 was expected since no orthologue was identified in the Drosophila genome by bioinformatic analyses (Gabaldón et al., 2005), and in Drosophila the N-terminus of NDUFC2 is displaced by the C-terminal extension of NDUFA11 (see Figure 1—figure supplement 5). Based on the same bioinformatic analyses, subunit NDUFA3 was also expected to be absent, but a matching subunit (Dmel gene CG9034) was detected by mass spectrometry in our preparation (see 'Materials and methods') and is clearly present in our density map in the location of mammalian-NDUFA3 in the membrane domain (see Figure 1—figure supplement 6). However, the sequence homology is weak and the structures of the two proteins diverge in the C-terminal membrane-extrinsic domain, with the obtuse-angled ‘turn’ that follows the TMH in the mammalian protein sterically blocked by the marginally extended C-terminal TMH of ND1 in Drosophila. NDUFA2, which has a characteristic thioredoxin fold and is widely conserved in eukaryotic complex I, is surprisingly absent from our Drosophila structure despite a highly conserved homologue in the Drosophila genome (Gabaldón et al., 2005). However, NDUFA2 interacts with only subunit NDUFS1 in the mammalian complex, and inspection of the (otherwise highly conserved) region of interaction in the Drosophila enzyme shows local disorder in a specific helix in the Drosophila NDUFS1 subunit (residues 673–684) that binds NDUFA2 in the mammalian enzyme. Although this result suggests NDUFA2 is associated with Drosophila complex I in vivo but has been lost during enzyme purification, detailed transcriptomic analyses (Brown et al., 2014; Leader et al., 2018) show that NDUFA2 (Drosophila ND-B8) expression is restricted principally to the male germline, and therefore the NDUFA2 protein is unlikely to be a constitutive component of complex I in somatic tissues.

Overall, Drosophila complex I is remarkably similar in its composition and structure to the mammalian enzyme, underlining expectations of the value of Drosophila as a model system for complex I research. Only further minor differences are present in some of the subunits (Figure 1—figure supplement 5). ND5-TMH1 is absent from the Drosophila subunit: although ND2, ND4, and ND5 have a canonical 14-TMH core structure, truncation of the N-terminal TMHs appears tolerated, consistent with them lacking specific catalytically active residues or features, and demonstrated by the 11-TMH form of subunit ND2 in bilateria that lacks the three N-terminal TMHs found in lower organisms (Birrell and Hirst, 2010). In addition, the structures of supernumerary subunits NDUFB6 and NDUFB1 are noticeably different in the Drosophila enzyme (Figure 1—figure supplement 5). Notably, all the substantial differences in the membrane domain (absence of NDUFC1 and ND5-TMH1, variations in NDUFA3, NDUFB6 and NDUFB1, extension of NDUFA11) are located on the ‘right’ side of the boot-shaped enzyme, perhaps because there is less evolutionary pressure on the right side than on the left, where interactions with complexes III and IV are central to the stabilisation of respiratory chain supercomplexes (Milenkovic et al., 2017).

Cryo-EM particle classification identified three distinct states in our preparation of Drosophila complex I, which we refer to as Dm1, Dm2, and Dm3 (Figure 1—figure supplement 2). The Dm1 class is the dominant class, containing ~60% of the particles, whereas the two minor classes, Dm2 and Dm3, each contain ~20%. On a global scale (Figure 2), the Dm2 state is ‘twisted’ relative to the Dm1 state: with the two models aligned on subunit ND1 in the ‘heel’ of the complex, the hydrophilic and membrane domains twist in opposite directions (there is no apparent opening or closing motion between the domains). A similar twisting relationship was identified between the active and deactive resting states of mammalian complex I (Zhu et al., 2016). However, standard biochemical assays used to detect the presence of the mammalian deactive state did not detect any deactive Drosophila enzyme, even after incubation at 37°C to promote deactivation (Figure 2—figure supplement 1), indicating that Dm2 is not directly comparable to the mammalian-type deactive state. In the mammalian deactive state, the equivalent residue to ND3-Cys41 (we use Drosophila numbering throughout) is exposed to solution and can be derivatised by N-ethylmaleimide (NEM), preventing reactivation of the deactive enzyme and its return to catalysis, whereas in the active state ND3-Cys41 is buried (Galkin et al., 2008). Our assays suggest that either ND3-Cys41 is buried and inaccessible to derivatisation in all three Dm1, Dm2, and Dm3 states, or that ND3-Cys41 is exposed in one or more state that is completely inactive, being unable to either reactivate or catalyse. For the Dm3 state, the most obvious global feature (Figure 2) is that the membrane domain appears ‘cracked’ at the interface between ND2 and ND4; the density for the adjacent subunit NDUFA11 is disordered, along with the adjacent N-terminus of NDUFS2 (Figure 2—figure supplement 2). These characteristics resemble those of the ‘slack’ state of bovine complex I (Chung et al., 2022b; Zhu et al., 2016), which is of uncertain biochemical and physiological relevance and which may result from destabilisation of the membrane-intrinsic domain following extraction from the membrane and delipidation by detergents during purification. To evaluate the three states of ‘resting’ Drosophila complex I further, we first focus on the Dm1 state and its relationship with known resting states of the mammalian enzyme.

In addition to differing in their global conformations, the mammalian active and deactive states are differentiated by the status of a set of local features in the core subunits (Agip et al., 2018; Blaza et al., 2018; Chung et al., 2022a; Chung et al., 2022b; Zhu et al., 2016). In the Dm1 state of Drosophila complex I, these features are all unambiguously in the active state (Figure 3a). First, ND6-TMH3 is clearly α-helical, it does not contain the π-bulge that is characteristic of the deactive state, and ND1-TMH4 is clearly in the ‘bent’ conformation of the active state (with Tyr149 pointing toward the E-channel), not the straight conformation of the deactive state (with Tyr149 pointing away from the E-channel). Second, the densities for the NDUFS2-β1−β2 loop that carries the His97 ligand to bound ubiquinone, the ND3-TMH1−2 loop that carries Cys41 (the biochemical marker of the mammalian active/deactive states), and the ND1-TMH5−6 loop, are all well-defined in the Dm1 density map and their conformations match the mammalian active-state conformations (they are not disordered as in the deactive state). ND3-Cys41, NDUFS2-His93, and ND1-Tyr134 meet in a trigonal junction at the top of ND1-TMH4 (Grba and Hirst, 2020), as they do in the active state (Figure 3c). Finally, the FRASPR motif in NDUFS7 that includes Arg119 matches its conformation in the mammalian active, not deactive, state.

Importantly, as expected from the ordered states of the loop structures that form the ubiquinone-binding site, the site is sealed and closed from the matrix (Figure 3b and c) as in the mammalian active state, not open to the matrix as in the deactive state (Agip et al., 2018; Blaza et al., 2018; Chung et al., 2022a; Chung et al., 2022b; Zhu et al., 2016). This observation indicates that the Dm1 state is a catalytically competent state, ready to bind and reduce the extended and hydrophobic ubiquinone-9 or ubiquinone-10 substrate (referred to as Q₉ for brevity). Indeed, density for Q₉ is observed within the site, although the Q₉ is only partially inserted, with its ubiquinone-headgroup in the central section of the channel, rather than ligated to the two proton-donor ligands NDUFS2-His97 and NDUFS2-Tyr146 as required for its reduction (Baradaran et al., 2013; Tocilescu et al., 2010). Partially inserted ubiquinones have been observed previously in several different species and states of complex I (Gu et al., 2022; Kampjut and Sazanov, 2020; Kravchuk et al., 2022; Parey et al., 2019; Soufari et al., 2020), but the headgroup typically sits slightly lower down the channel than observed here, in an array of positions distributed largely around the hydrophilic ‘kink’ of the channel (Figure 3—figure supplement 1). The Drosophila Q₉ headgroup is bound 11 and 14 Å away, respectively, from its proposed ligating partners NDUFS2-His97 and NDUFS2-Tyr146, between the ‘1^F’ site described in Sus scrofa complex I (Gu et al., 2022) and the ‘Q_m’ site described in Escherichia coli (Kravchuk et al., 2022). The wide spectrum of headgroup positions identified in different complex I structures suggests that substrates may shuttle in a step-wise manner, occupying numerous sites of localised energy minima (Chung et al., 2021; Hoias Teixeira and Menegon Arantes, 2019; Warnau et al., 2018). Thus, although the site observed here is clearly separated from the reactive site, it is only broadly defined and not a highly specific site.

In the mammalian complex, two subunits, NDUFA5 on the hydrophilic domain and NDUFA10 on the membrane domain, meet in the corner of the L-shape forming an interface between the two domains. Upon deactivation of the mammalian enzyme, the altered disposition of the hydrophilic and membrane domains changes the NDUFA5/NDUFA10 interface and decreases their contact area (Agip et al., 2018; Zhu et al., 2016). The nature and extent of the interface thereby provide an easy way to evaluate the active/deactive status of mammalian structures. Both subunits are present in the Drosophila enzyme, and Figure 4a compares their relative positions in the Dm1 active state to their relative positions in the active and deactive mammalian states. With the structures aligned to subunit NDUFA10, subunit NDUFA5, which is dominated by a three-helix bundle, clearly lies in an intermediate position in the Dm1 state, it does not overlay its position in the mammalian active state. The N-terminus of hydrophilic-domain subunit NDUFS2 that runs along the top of the membrane domain is also in an intermediate position. However, the NDUFA5/NDUFA10 contact area still matches closely to that observed in the mammalian active state (388 Ų vs. 354 and 131 Ų in the bovine active and deactive states [Chung et al., 2022b], respectively), and the NDUFA5/NDUFA10 interface is clearly different as it incorporates the N-terminus of the NDUFS4 subunit (residues 33–49; Figure 4b), which is substantially extended relative to in mammalian species (Figure 4c). Contacts between the NDUFS4 N-terminal ‘tether’ and subunits NDUFA5 and NDUFA10 of 182 and 427 Ų, respectively, further stabilise the interface and, by extension, the relative disposition of the hydrophilic and membrane domains in the Drosophila Dm1 active state.

Further comparison of the mammalian and Drosophila active-state structures showed that subunits NDUFS2, NDUFS7, and ND1 that constitute the ubiquinone-binding site (Figure 3), overlay closely (RMSD 0.574 between Dm1 and the bovine active state [PDB ID: 7QSK]; Chung et al., 2022b) but that the structures then diverge along the membrane domain, shifting the position and orientation of subunit ND2 (Figure 5a). As NDUFA5 is bound to NDUFS2 and NDUFA10 to ND2, their relative positions thus also change. Within the connecting subdomain between ND1 and ND2 that contains subunits ND3, ND6, and ND4L, the arrangement of the TMHs in subunit ND6 (Figure 5b) clearly differs between the Drosophila and mammalian enzymes (despite them both containing a fully α-helical ND6-TMH3). In particular, ND6-TMH4 is markedly displaced, enhancing a cleft between ND6 and ND1 in which three phospholipid molecules are observed (Figure 1—figure supplement 4). The position of ND6-TMH4 is remarkably variable between different species and states of complex I, suggesting that it is not functionally important (Figure 5—figure supplement 1). Structures around ND6-TMH4 also vary (Figure 5b and c): (i) the ND6-TMH4–5 loop is restructured and its β-hairpin disrupted by neighbouring ND4L-TMH1, (ii) ND6-TMH1 is displaced away from TMH4 to avoid steric clashes, (iii) the ND6-TMH3–4 loop is restructured to accommodate the movement of TMH4, and (iv) the C-terminal loop of NDUFA9, located just above the reordered ND6-TMH3–4 loop, is retracted. The ND3-TMH1–2 loop, which is also adjacent to the restructured region, remains ordered with ND3-Cys41 occluded (Figure 5c). Strikingly, the residues of the central axis that link the terminus of the E-channel to the start of the first antiporter-like subunit (ND2) are not affected by the altered connecting subdomain structure, which thus adjusts the relative disposition of subunits in the hydrophilic and membrane domains in the Drosophila Dm1 state without affecting the catalytic machinery of the active state structure.

Inspection of the set of local features in the core subunits (Agip et al., 2018; Blaza et al., 2018; Chung et al., 2022a; Chung et al., 2022b; Letts et al., 2019; Zhu et al., 2016) that differentiate the mammalian active and deactive states in the two minor states revealed that both the Dm2 and Dm3 states also most closely correspond to the mammalian active state. Figure 6a shows that, in the Dm2 state, two features of the mammalian deactive state are present: a π-bulge has formed in ND6-TMH3 and ND1-TMH4-Tyr149 has ‘flipped’ its conformation (Chung et al., 2022b; Grba and Hirst, 2020; Kampjut and Sazanov, 2020). However, all the other key elements remain in their active states, including ND1-TMH4, which remains in its bent conformation, and as a result the ubiquinone-binding site remains enclosed and sealed from the matrix (Figure 6b and c). The same is true for the Dm3 state (Figure 6—figure supplement 1). Notably, the trigonal junction between ND3-Cys41, NDUFS2-His93, and ND1-Tyr134 (Grba and Hirst, 2020) is preserved in all three states, occluding the Cys from the matrix and explaining why a mammalian-type deactive state of Drosophila complex I could not be trapped in biochemical assays (Figure 2—figure supplement 1). We conclude that, in contrast to the mammalian enzyme, Drosophila complex I does not form an ‘open’ resting state, it rests only in ‘closed’ conformations with the ubiquinone-binding site enclosed and sealed from the matrix.

Further comparison of the Dm1 and Dm2 states (Figure 7a and b) shows that the ‘twisting’ motion that relates their global conformations originates in changes in the ‘connecting subdomain’ between the rigid and mobile subdomains described in Figure 5a, where the deactive-like elements of Dm2 are located (the π-bulge in ND6-TMH3 and ND1-TMH4-Tyr149). The twisting motion displaces the N-terminus of subunit NDUFS2 (adjacent to ND6-TMH5) and changes the NDUFA5/NDUFA10 interface, displacing and disordering the N-terminus of subunit NDUFS4 (Figure 7c) and causing a small decrease in the interface area (from 388 Ų in Dm1 to 333 Ų in Dm2, relative to 354 and 131 Ų in the active and deactive states of bovine complex I; Chung et al., 2022b). Inspection of the region around the π-bulge in ND6-TMH3 revealed a further striking change between the Dm1 and Dm2 states. In Dm1, the tail of a phosphatidylcholine molecule is intercalated into the structure, sterically obstructing the rotation of bulky residues on ND6-TMH3 around the helical axis to form the π-bulge (Figure 7a and b). It is absent from the Dm2 state (and also from Dm3, where the local protein conformation matches Dm2; Figure 6—figure supplement 1). In the Dm1 state, the phosphatidylcholine headgroup stabilises the ND6-TMH3–4 loop at the top of ND6-TMH4, whereas its absence in Dm2 allows the TMH3–4 loop and TMH4 to move, along with a further adjustment to the adjacent C-terminal loop of NDUFA9 and displacement of ND3-TMH2–3. The lipid may either have been ejected during relaxation of ND6-TMH3 into a π-bulge structure, or removed during detergent extraction, promoting π-bulge formation.

Flies of a common wild type-equivalent genotype, isogenic w¹¹¹⁸ (RRID:BDSC_6326), were obtained from Bloomington Drosophila Stock Center (RRID:SCR_006457), raised and kept under standard conditions in a temperature-controlled incubator with a 12 hr:12 hr light:dark cycle at 25°C and 65% relative humidity, on food consisting of agar, cornmeal, molasses, propionic acid, and yeast. Approximately 5500 mixed adults collected 5 days after eclosion were used for the preparation of the cryo-EM sample.

All experimental procedures were carried at 4°C unless otherwise stated. One volume of w¹¹¹⁸ flies (1 mL is equivalent to ~100 flies) was mixed with five volumes of homogenisation buffer containing 20 mM Tris-HCl pH 7.8, 250 mM sucrose, 2 mM EDTA, 2 mM EGTA, 1% (w/v) fatty acid-free bovine serum albumin (BSA, Merck) and 1× EDTA-free cOmplete protease inhibitor cocktail (Roche) (one tablet per 50 mL of buffer). Mitochondria were prepared using a differential centrifugation method. Briefly, 20 mL of fly suspension were homogenised by 10 strokes with a motor-driven Teflon pestle at 1300 rpm in a 30 mL Wheaton glass homogeniser. The homogenate was then centrifuged at 1000 × g for 5 min, and the supernatant filtered through a muslin cloth to remove cuticles. The same process was repeated a second time. Then, mitochondria were pelleted at 3000 × g for 10 min and washed with 5 mL of homogenisation buffer but without BSA. Finally, mitochondria were collected by centrifugation at 7000 × g for 10 min and resuspended in 5.8 mL (56 mg of protein) of resuspension buffer containing 20 mM Tris-HCl (pH 7.8), 20% glycerol (v/v), 2 mM EDTA, 2 mM EGTA, 1% and 1× EDTA-free cOmplete protease inhibitor cocktail (one tablet per 50 mL of buffer). Isolated mitochondria were stored at –80°C until further use. Mitochondrial membranes were prepared as described previously for mouse samples (Agip et al., 2018). Defrosted mitochondria were diluted to 5 mg mL^–1 in resuspension buffer then ruptured on ice with a Q700 Sonicator (Qsonica) in three intervals (5 s bursts each followed by a 30 s pause) at an amplitude setting of 65%. Membranes (38 mg of protein) were collected by centrifugation at 75,000 × g for 1 hr, then resuspended to 4.9 mg mL^–1 in the same buffer and stored at –80°C.

Purification of Drosophila melanogaster complex I followed the same procedure as previously described for mouse complex I, with minor adjustments (Agip et al., 2018). While being continuously stirred on ice, mitochondrial membranes (7.6 mL at 4.9 mg mL^–1) were solubilised for 30 min by the drop-wise addition of dodecyl-β-D-maltoside (DDM) to a final concentration of 0.75% from a 10% stock solution. The solubilised membranes were then centrifuged at 48,000 × g for 30 min and the clarified supernatant loaded on to a Hi-Trap Q HP anion exchange column (1 mL; Cytiva) pre-equilibrated with elution buffer A (20 mM Tris-HCl pH 7.8°C, 2 mM EDTA, 2 mM EGTA, 0.1% DDM, 10% ethylene glycol [v/v, VWR], 0.005% asolectin [Avanti], and 0.005% CHAPS [Calbiochem]) and operated at a flow rate of 0.3 mL min^–1. The column was washed with several column volumes of buffer A until the 280 nm absorbance reached the baseline. Unwanted proteins were eluted with seven column volumes of 20% buffer B (buffer A + 1 M NaCl), then complex I was eluted with an additional seven column volumes of 35% buffer B. Fractions containing complex I (ca. 3 mL) were pooled and concentrated to ca. 100 µL using an Amicon-Ultra filter device (100 kDa molecular weight cut-off; Amicon, Millipore). The concentrated sample was then injected onto a Superose 6 Increase size exclusion column (150 × 5 mm; Cytiva) pre-equilibrated in buffer C (20 mM Tris-HCl pH 7.8°C, 150 mM NaCl, and 0.05% DDM) operated at a flow rate of 0.03 mL min^–1. All chromatographic procedures described were carried out using an ÄKTA micro FPLC system (Cytiva) with elution monitored at 280 and 420 nm. Complex I concentrations were estimated at 280 nm (ε = 0.2 mg mL^–1 mm^–1). The total collected protein was estimated at 0.6 mg, and the peak concentration was 3.4 mg mL^–1.

All activity measurements were measured on a 96-well Spectramax 384 plate reader at 32°C. For NADH:decylubiquinone (dQ) oxidoreductase activities, NADH (200 µM final concentration) was used to initiate catalysis by complex I (0.2 µg mL^–1) with 200 µM dQ, 0.15% (w/v) asolectin, and 0.15% (w/v) CHAPS in 20 mM Tris-HCl (pH 7.55). NADH oxidation was monitored at 340–380 nm (ε = 4.81 mM^–1 cm^–1), and was confirmed to be sensitive to rotenone and piericidin A. The cryo-EM sample had an activity of 7.3 ± 0.3 µmol min^–1 mg^–1 (mean ± SD; n = 4).

For evaluation of the active/deactive state ratio of Drosophila complex I using the N-ethylmaleimide (NEM) assay (Galkin et al., 2008; Yin et al., 2021), 4 mg mL^–1 mitochondria were incubated with 2 mM NEM or the equivalent volume of DMSO on ice for 20 min., before determining the NADH:O₂ oxidoreductase activity. The mitochondria had been frozen for storage before measurement. To attempt to deactivate the complex, the mitochondria were incubated at 37°C for 30 min (equivalent to, or longer than, the treatments used to deactivate complex I in mammalian mitochondrial membranes [Agip et al., 2018; Blaza et al., 2018]). NADH:O₂ oxidoreductase activities were measured in 20 mM Tris-HCl (pH 7.55) using 10 µg mL^–1 mitochondria and 10 µg mL^–1 alamethicin, and initiated using 200 µM NADH. NADH oxidation was monitored as described above.

UltrAuFoil gold grids (0.6/1, Quantifoil) (Russo and Passmore, 2014) were prepared for Drosophila complex I as described previously (Blaza et al., 2018). First, the grids were glow discharged with plasma under vacuum for 90 s at 20 mA then incubated for 7 days under anaerobic and room temperature conditions in a solution of 5 mM 11-mercaptoundecyl hexaethylene glycol (SPT-0011P6, SensoPath Technologies) (Meyerson et al., 2015). Grids were then washed several times in ethanol and left to dry. Complex I (3.4 mg mL^–1) was then applied (3 µL per grid) to the treated grids in a Vitrobot Mark IV (Thermo Fisher Scientific) set to 4°C and 100% relative humidity. Grids were blotted for 10 s with a force setting of –10, before being plunged into liquid ethane. Frozen grids were then stored in liquid nitrogen before screening and data collection.

Both cryo-EM screening and high-resolution image collection were carried out on a Titan Krios (Thermo Fisher Scientific) at University of Cambridge cryo-EM facility. The Titan Krios microscope for data collection was operating at an accelerated voltage of 300 kV and equipped with a Gatan K2 detector utilising a GIF quantum energy filter with a slit width of 20 eV. The microscope was operated in electron counting mode with a nominal sampling rate of 1.07 Å pix^–1 (nominal magnification of 130,000) and a dose rate of ca. 4.18 electrons Å^–2 s^–1. The specimen was radiated for 10 s over 40 frames with a total exposure amounting to ca. 42 electrons Å^–2. A 100 µm and 50 µm objective and C2 aperture, respectively, were inserted during high-resolution imaging. The microscope was operated with EPU software and the defocus range was set to –1.0 to –2.0 µm, with an autofocus routine run every 5 µm.

All 3082 collected movies were subjected to processing by RELION-3.0 and 3.1 (Zivanov et al., 2020; Zivanov et al., 2018) except where stated otherwise. Micrographs were motion-corrected using MotionCor2 (Zheng et al., 2017) with 5 × 5 patches, and contrast transfer function (CTF) parameters estimated using CTFFIND-4.1 (Rohou and Grigorieff, 2015) in RELION-3.0. In parallel, the motion-corrected micrographs were exported and subjected to particle autopicking using a general model without training in crYOLO 1.5.3 (Wagner et al., 2019), resulting in 194,538 picked particles. Ice-contaminated micrographs were removed to give 180,342 particles from 2852 micrographs. Particles were extracted with an initial downscaling to 6.0 Å pixel^–1 (box size of 80) and subjected to initial 2D and 3D classification steps to remove junk particles, yielding a total of 93,332 particles. These particles were re-extracted at the nominal pixel size of 1.07 Å pixel^–1 (box size of 450) and used to reconstruct a 3.74 Å resolution map using the 3D autorefinement procedure in RELION, at the calibrated pixel size of 1.048 Å pixel^–1 (Spikes et al., 2020). The active state map of mouse complex I (EMD-4345) (Agip et al., 2018) was used as a reference map for the 3D reconstruction. Bayesian polishing (Zivanov et al., 2019) was then applied and CTF parameters, including astigmatism, defocus, and beam tilt, estimated using the CTF refinement procedure in RELION-3.0. Particles were subjected to additional rounds of classifications, to further remove junk and bad complex I particles. From hereon, all data processing was performed in RELION-3.1, at the nominal pixel size of 1.07 Å pixel^–1, then corrected to the calibrated pixel size of 1.048 Å pixel^–1 at the postprocessing or local resolution stages. The particles were subject to iterative rounds of CTF refinement (Zivanov et al., 2020), to estimate anisotropic magnification, beam tilt, trefoil, fourth-order aberration, and per-particle defocus, astigmatism and B-factor parameters. Particles with an rlnNrOfSignificantSamples value greater than 3000 were removed to give 65,864 particles. Using a complex I mask (generated from a working model using RELION MaskCreate) and with solvent flattening, the global resolution of the 3D refined map was 3.23 Å (according to a gold-standard Fourier shell correlation [FSC] of 0.143; Rosenthal and Henderson, 2003). 3D classification (number of classes, K = 5, local angular search to 0.2° sampling) was then performed, and three complex I classes, Dm1, Dm2, and Dm3, were identified and retained, containing 37,608, 12,343, and 13,520 particles, respectively, a ratio of roughly 3:1:1. Clear cryo-EM densities for Dm2-specific local features, including the ‘flipped’ ND1-TMH4-Tyr149 and the ND6-TMH3 π-bulge, revealed no evidence for Dm1 contamination in the Dm2 population. The ‘Focus-Revert-Classify’ classification strategy (Letts et al., 2019), applied using the regularisation parameter t = 8 and K = 5, yielded comparable population distributions (three complex I classes matching Dm1, Dm2, and Dm3, plus two junk classes) whilst 3D classification without alignment using t = 20 and K ≤ 12 yielded two <4 Å complex I classes, with the major class matching Dm1 and the minor class an apparent mixture of Dm2 and Dm3. The 3D classification approach with local angular sampling was therefore employed to give the final set of Dm1, Dm2, and Dm3 particles as described above. Using model-generated (Dm1 and Dm2) or map-generated (Dm3) masks and solvent-flattening, the three classes refined to 3.28, 3.68, and 3.96 Å resolution, respectively. Global resolutions were estimated from two independent half maps using a gold-standard FSC of 0.143 (Rosenthal and Henderson, 2003) in RELION postprocess. The final map was globally sharpened (or blurred) in RELION postprocess using user-provided B-factor values. The model-generated or map-generated mask used for 3D refinement procedures and resolution estimation was generated in UCSF ChimeraX (Pettersen et al., 2021) using the molmap (Dm1 and Dm2) or vop threshold (Dm3) functions, before being low-pass filtered to 15 Å and having a 6-pixel soft cosine edge added using RELION MaskCreate. Local resolution was estimated using RELION LocRes. Mollweide projections were plotted using Python and Matplotlib, and the degree of directional resolution anisotropy calculated using the 3DFSC program suite (Tan et al., 2017).

Model coordinates were built into the 3.3 Å resolution Dm1 Drosophila complex I map, with a published mouse complex I structure (PDB ID: 6G2J) (Agip et al., 2018) serving as a homology model. SWISS-MODEL (Waterhouse et al., 2018) was used to generate an initial model for each subunit and the models rigid-body fitted into the map using Chimera (Pettersen et al., 2004). MODELLER (Webb and Sali, 2016) was used to generate models for subunits NDUFB6, NDUFB9, and NDUFB8 as the homology models were unsatisfactory. The sequence for the Drosophila NDUFA3 subunit (UniProt ID: Q9W380; Dmel gene CG9034; FlyBase ID: FBgn0040931) was identified in routine peptide mass spectrometry analyses of the purified enzyme. The evidence for Q9W380 relies on a single peptide (LGYVVYR, 9.1% coverage) but the MASCOT score is 43, well above 30 (the 99% confidence limit), and inspection of the distribution of tryptic cleavage sites suggests no further peptides would be expected to be detected. Models for subunits NDUFA2 and NDUFC1 were deleted due to the lack of corresponding densities in the cryo-EM maps, and N- and C-terminal extensions were built where necessary. It was noted that densities for the N-termini of ND1 and ND5 were extended beyond the reviewed UniProt sequences (P18929 and P18932, respectively). Therefore, to incorporate the correct translation start site, UniProt IDs C7DZL9 and C7DZL4 (Stewart and Beckenbach, 2009) were used to build models for subunits ND1 and ND5, respectively. UniProt ID A0A024E3A5 was used for subunit NDUFA9 to include a L174F mutation supported by the Drosophila cryo-EM density features. Notably, the N-terminus of core subunit NDUFS7 is resolved for the first time in Drosophila complex I (Figure 1—figure supplement 5). Preliminary model building and real-space refinements were carried out in Coot 0.9-pre (Casañal et al., 2020). Then GPU-powered ISOLDE 1.0 (Croll, 2018), which implements a molecular dynamic approach to model refinement, was used to iterate through the model, improving the map-to-model fit, resolving clashes and maintaining good protein stereochemistry. Emerging modelling errors were monitored using a real-time validation functionality present in ISOLDE and corrected. Densities for existing and additional phospholipid molecules were identified with the Unmodelled blobs tool in Coot 0.9.6.2 (Casañal et al., 2020). All non-cardiolipin phospholipids were modelled as phosphatidylethanolamines, the largest component of the phospholipid composition of the Drosophila mitochondrial membranes (Jones et al., 1992), unless density features indicated phosphatidylcholine to be more likely. Lipid tails were clipped where necessary using the delete tools in Coot and PyMOL 2.5.2 (Schrodinger, 2022). dGTP was modelled in subunit NDUFA10 (Molina-Granada et al., 2022). The model was then Curlew all-atom-refined using Coot and real-space refined against the active-state map using phenix.real_space_refine in Phenix 1.18.2–3874 (Liebschner et al., 2019) with custom geometry restraints. Ligand restraints were generated using Phenix eLBOW. No secondary structure restraints were used during real-space refinement of the Dm1 model. The model was checked manually in Coot, new resolvable regions built, and rotameric and/or Ramachandran outliers corrected. Atom resolvabilities (Q-scores) were calculated using MapQ (Pintilie et al., 2020) and any persisting outliers identified and corrected. The model was then real-space refined in Phenix as described above to produce the final Dm1 model.

To build the Dm2 model, the Dm1 Drosophila model was rigid-body fitted into the Dm2 map using the Fit in map tool in UCSF ChimeraX (Pettersen et al., 2021) followed by rigid-body fitting by subunit in Phenix 1.18.2–3874, and Curlew all-atom-refined using Coot 0.9.6.2. The Dm2 model was manually inspected and new resolvable regions, less resolved regions, and/or conformationally different regions built or deleted manually in Coot, and locally refined in ISOLDE 1.4 (Croll, 2018). Q-scores were calculated using MapQ, and any outliers identified and corrected. Existing lipids were checked against their densities and deleted where appropriate; lipid tails were similarly clipped where necessary as described above. The model was then real-space refined against the Dm2 map in Phenix 1.18.2–3874 with custom geometry restraints and secondary structure restraints (identified by ksdssp). Iteratively, rotameric and Ramachandran outliers were corrected manually in Coot and real-space refined in Phenix. The final real-space refinement for the Dm2 model was performed without secondary structure restraints in Phenix. The model statistics for the Dm1 and Dm2 classes (Table 1) were produced by Phenix, MolProbity (Chen et al., 2010), and EMRinger (Barad et al., 2015). Model-to-map FSC curves were generated using phenix.validation_cryoem in Phenix.

Individual subunits from the Dm2 model were rigid-body fitted into the Dm3 map in UCSF ChimeraX (Pettersen et al., 2021) to generate a tentative model for the Dm3 state for visualisation.

RMSD calculations between models were performed using the Align command in PyMOL 2.5.2 (Schrodinger, 2022). Buried surface area between subunits were calculated using the measure buriedarea command in UCSF ChimeraX (Pettersen et al., 2021). The interior surface of the ubiquinone-binding channel was predicted using CASTp (Tian et al., 2018), which computes a protein surface topology from a PDB model. The default 1.4 Å radius probe was used and the results were visualised in PyMOL using the CASTpyMOL 3.1 plugin and by UCSF ChimeraX.