Cryo-EM structures reveal intricate Fe-S cluster arrangement and charging in Rhodobacter capsulatus formate dehydrogenase

Metal-containing formate dehydrogenases (FDH) catalyse the reversible oxidation of formate to carbon dioxide at their molybdenum or tungsten active site. They display a diverse subunit and cofactor composition, but structural information on these enzymes is limited. Here we report the cryo-electron microscopic structures of the soluble Rhodobacter capsulatus FDH (RcFDH) as isolated and in the presence of reduced nicotinamide adenine dinucleotide (NADH). RcFDH assembles into a 360 kDa dimer of heterotetramers revealing a putative interconnection of electron pathway chains. In the presence of NADH, the RcFDH structure shows charging of cofactors, indicative of an increased electron load.

N ature provides a wealth of enzymes that link the two halfreactions of a redox reaction through an electron transfer pathway. Biological processes, such as respiration, anaerobic metabolism and nitrogen fixation, depend on these enzymes and offer inspirational alternatives for difficult chemical conversions. One of these oxidoreductases, the metal-containing formate dehydrogenase, catalyses the following redox reaction: In standard state, the equilibrium of the reaction favours formate oxidation at the active site Mo or W atom, which is energetically coupled to the reduction of oxidised nicotinamide adenine dinucleotide (NAD + ) in cytoplasmic FDHs e.g. from Rhodobacter capsulatus and Cupriavidus necator. In the presence of an excess of reducing equivalents and under physiological, cellular conditions the reaction is reversible, as shown for several FDHs and formyl-methanofuran dehydrogenases [1][2][3][4][5][6][7][8] . This family of enzymes has also been termed CO 2 reductases 9 . Enzymatic reduction of carbon dioxide to formate would allow for storage of hydrogen as a fuel for industrial applications 2,9,10 as well as carbon sequestration from the atmosphere 11,12 , making these enzymes interesting targets for biotechnological applications. Bacterial FDHs can be divided into different classes, which are distinguished by their subcellular localisation, subunit organisation and cofactor composition 13 . They all share a bis-metalbinding pterin (molybdenum or tungsten) guanine dinucleotide (bis-MGD) cofactor-containing subunit that also binds a proximal [4Fe-4S] cluster, an arrangement that is conserved in other enzymes, e.g. bacterial nitrate reductases 14 . The majority of mesophilic, prokaryotic FDHs coordinate Mo as active site metal. This Mo atom is ligated to two pterin dithiolenes, a sulphur atom and either a cysteine or a selenocysteine. FDHs containing the latter ligand were found to be rather oxygen sensitive 13,15 . All FDHs contain two additional highly conserved residues in the active site, a histidine and an arginine 16,17 . So far, no structural information is available on cytoplasmic, NAD + -dependent formate dehydrogenases.
Here, we present the 3.3 Å cryo-EM structures of RcFDH as isolated and in the presence of NADH. The structures reveal a complex arrangement of Fe-S clusters in the dimer, a conserved binding mode of the FdsD to the FdsA subunit and that NADH reduction leads to charging of electron carrying cofactors of RcFDH.

Results
Domain architecture of RcFDH. We expressed the RcFDH operon fdsGBACD homologously in R. capsulatus cells and purified the complex aerobically in the presence of 10 mM azide, a potent inhibitor of FDH ( Supplementary Fig. 1a) 18 . The purified complex consists of four subunits, FdsA, FdsB, FdsG and FdsD (Supplementary Figure 1b). In contrast to recent reports on the heterologously expressed R. capsulatus enzyme and the FDH from C. necator, FdsD is retained as a subunit in the active enzyme ( Supplementary Fig. 1c) 6,16,19 . Earlier analyses also support a FdsABGD complex arrangement 20,21 , indicating that likely all NAD + -dependent FDHs have a similar overall subunit composition. The homologously expressed R. capsulatus enzyme has a similar cofactor saturation and overall activity as compared to the previously reported enzyme expressed in Escherichia coli 6 .
To solve the structure of the functional RcFDH complex, we subjected the as isolated, azide inhibited sample to cryo-electron microscopy (cryo-EM) and single particle analysis (Supplementary Fig. 2). Two-fold symmetry of the particle can distinctively be derived from reference-free 2D class averages. Subsequently, we generated a 3D reconstruction with imposed C2 symmetry yielding a final reconstruction with an overall resolution of 3.3 Å as determined by gold-standard Fourier shell correlation (FSC) at 0.143.
The RcFDH complex forms a 360 kDa dimer of FdsABGD heterotetramers (Fig. 1). The heterotetramers adopt an elongated structure with dimensions of 140 ×80 ×77 Å and are arranged in an almost perpendicular back-to-back orientation in the dimer. The dimer interface comprises an area of 1369 Å 2 and is solely formed by the two FdsA (105 kDa, 958 amino acids) subunits. The diaphorase unit consisting of FdsB (55 kDa, 500 amino acids) and FdsG ( (Fig. 2b). This cluster was suggested to minimise reactive oxygen species (ROS) production due to flavin semiquinone radicals in the presence of oxygen by temporarily storing electrons 26 , inducing structural rearrangements that stabilise NAD + binding upon reduction 28 , or increasing overall enzyme stability 29 . Although the diaphorase structure of RcFDH shares high similarity with those of other oxidoreductases, the amino-terminal ferredoxin-like domain of FdsB is absent in related NADH-quinone oxidoreductases and the NAD + -reducing [NiFe] hydrogenases ( Supplementary Fig. 3).
FdsD has been shown to positively influence the insertion of bis-MGD into RcFDH 6 and was predicted to either function as a chaperone for bis-MGD insertion or to stabilise the quaternary structure of FdsA. Homologues of FdsD can only be found in NAD + -dependent formate dehydrogenases, but not in other formate dehydrogenases or formylmethanofuran dehydrogenases. Up to now, no structural data are available on FdsD. Our cryo-EM map reveals that FdsD folds into a four-helix bundle resembling domain 1 of methionine synthase 30 , albeit no apparent sequence homology to known proteins was identified. FdsD interfaces with 23 out of its 69 amino acids with both MGD coordinating domains of FdsA and with a loop that extends from the C-terminal cap domain shielding the bis-MGD cofactor (Fig. 2c). Interaction is mediated by several hydrogen bonds, by three salt bridges, and by aromatic stacking between Trp 917 in FdsA and two conserved phenylalanines (Phe 16 , Phe 17 ) in FdsD (Fig. 2c, Supplementary Fig. 4a). Trp 917 is located at the tip of the cap loop of FdsA and is highly conserved in all FDH complexes predicted to contain FdsD (Fig. 2d). Both ends of the loop contain conserved aromatic and charged residues that interact with both MGD cofactors of FdsA. We hypothesise that the cap loop functions as a sensor for bis-MGD insertion during FDH assembly. FdsD binding to FdsA locks the cap domain of FdsA in place and might prevent damage or loss of the cofactors.
The Mo active site of RcFDH. The bis-MGD containing active site of as isolated RcFDH structurally resembles that of oxidised FdhF (Fig. 3a, Supplementary Movie 1). The EM map around molybdenum indicates six ligands coordinated in a trigonal prism geometry. The rectangular base of the prism is formed by the two dithiolene groups of the bis-MGD molecule coordinating molybdenum from one side. The two remaining coordination sites are occupied by the Cys 386 sulphur and a small ligand that is oriented towards Val 592 . As evidenced by the full reduction of the bis-MGD containing enzyme with formate ( Supplementary  Fig. 1c), a terminal sulfido ligand likely occupies this site 18 . The active site residues Arg 587 and His 387 are proposed to position formate for C-H bond cleavage and to elevate the pK a of the cysteine ligand, respectively 15 . The side chain orientation of both residues and the coordination of molybdenum largely resembles the arrangement observed in oxidised FdhF (PBD-ID 1fdo [https://doi.org/10.2210/pdb1FDO/pdb]) indicating that the active site molybdenum of RcFDH is present in the oxidised state. There is no clear evidence for stochiometric binding of azide in any particular location of the EM map. Conversely, azide inhibition of RcFDH under the conditions used in cryo EM is not mediated by stochiometric, direct binding to molybdenum.
Two tunnels can be identified in RcFDH starting at the molybdenum and separating at the active site residue Arg 587 into different exits (Fig. 3b). The pore of the shorter tunnel is mainly formed by polar and charged residues suggesting channelling of hydrophilic substrate from an entry site near FdsD to the active site. The tunnel forming residues show high conservation to those in oxidised FdhF (Supplementary Table 2). The second tunnel bears predominantly hydrophobic residues suggesting the possibility of gas transport to or away from the active site. FdhF contains a similar hydrophobic channel, which is however blocked by Val 145 and Met 157 in place of glycine residues at these position in RcFDH ( Supplementary Fig. 4b, Supplementary Table 3). Albeit the path of the hydrophobic tunnel differs from the CO 2 tunnel proposed for formyl-methanofuran dehydrogenase 3 , the active site residue Arg 587 might analogously control gate opening to each channel facilitating efficient catalysis. Intriguingly, the glycine residues are conserved in NAD + -dependent FDHs while other FDHs display larger hydrophobic residues at this position effectively occluding the tunnel ( Supplementary  Fig. 4c).
Cofactor arrangement in RcFDH. The direct electron transfer chain between bis-MGD and FMN of the heterotetramer measures 76 Å and consists of five Fe-S clusters: A1, A2, A3, A5 and B6 (Fig. 2a). All edge-to-edge electron transfer distances account for <14 Å and are thus within a reasonable distance for physiological electron transfer 31 . The high structural similarity between TtRC I and RcFDH is reflected in a good match of the Fe-S cluster positioning, when the structures are superimposed (Fig. 2b). The remote cluster N7 in complex I corresponds to cluster A1 in RcFDH, which can receive an electron from the reduced molybdenum atom. While cluster N7 is disconnected from other Fe-S clusters and is regarded as an evolutionary remnant in T. thermophilus, cluster A2 couples cluster A1 with the remaining electron transfer chain in RcFDH. The existence of a [4Fe-4S] cluster in a position equivalent to A2 in respiratory complex I of Campylobacter jejuni, Helicobacter pylori and A. aeolicus was predicted from sequence analysis 32    bonding distance, but the EM map suggests they are essentially coordinated to their corresponding cluster ( Supplementary  Fig. 4d). The FdsA subunit interface is formed by several hydrogen bonds and hydrophobic interactions. In particular, Cys 121 is flanked by two intercalating leucine residues (Leu 119 , Leu 122 ), which are conserved in NAD-dependent FDHs, but not in complex I ( Supplementary Fig. 4e). Leu 122 , which has the largest buried surface area in the interface, is more conserved than Leu 119 . The four Fe-S clusters A3-A4-A4′-A3′, lie all within a similar electron transfer distance (Fig. 2a) and potentially allow for electron transfer between the two FDH protomers. For membrane-bound [NiFe] hydrogenase I of E. coli it has been shown that electron transfer between protomers in the quaternary structure is important for recovery of the active site after O 2 attack 36,37 . This mechanism contributes to the oxygen tolerance of [NiFe] hydrogenases. The concrete nature of cluster A4's function will be the subject of future studies.
NADH-reduced structure of RcFDH. Our EPR spectroscopic characterisation shows that NADH treatment results in partial generation of the paramagnetic Mo V oxidation and respective reduction of the Fe-S clusters, without generation of an FMN • radical ( Supplementary Fig. 6 and Supplementary Table 4). The Mo V and detected Fe-S clusters bear resemblance to the FDHs characterised from Methanobacterium formicicum and Methylosinus trichosporium, and more recently from Cupriavidus necator 21,34,38 . Qualitatively, respective measurements show that these cofactors are partially reduced following NADH treatment ( Supplementary Fig. 5a, b). The nonquantitative integrated spin concentration at 12 K (reflecting weakly power-saturated [4Fe-4S] clusters) and 80 K (reflecting slow-relaxing Fe-S clusters and Mo V ) and the decreased spin concentration relative to the stronger reductant sodium dithionite support this assessment (Supplementary Table 4), in addition to the comparable partial reduction of FMN and Fe-S clusters by NADH observed at NADH:FDH ratios of 4000 and 20 as recorded by UV-visible spectroscopy ( Supplementary Fig. 5c, d). EPR spectroscopy and UV-Vis reduction spectra collectively reflect an enzyme that has underwent incomplete reduction by NADH. This behaviour is consistent with previous reports on other molybdoenzymes like xanthine dehydrogenase 39 . The inability of NADH to completely reduce the enzyme might be dependent on the redox potentials of the cofactors. Furthermore, the presence of highly inhibitory concentrations (10 mM) of azide might prevent complete reduction of the enzyme.
In order to structurally investigate this partially reduced state of the enzyme, we determined the cryo-EM structure of RcFDH in the presence of NADH and azide at 3.2 Å resolution ( Supplementary Fig. 6). The atomic model derived from this structure overlays with the coordinates of the as isolated state with a root mean square deviation of 0.31 Å over all but 20 Cα atoms of the four peptide chains (Supplementary Fig. 7, Supplementary Table 1). The strong similarity between both structures extends to the active site residues and cofactors, whose location and orientation are indistinguishable between the two states (Supplementary Movie 2). None of the structures shows dissociation of Cys 386 from the active site Mo. Also, no structural changes were observed at the pyranopterins of the bis-MGD likely reflecting that the cofactor has not been reduced to the Mo IV state 40 .
The difference density between the maps of the NADH reduced and the as isolated enzyme shows a distinct density for NADH near the FMN binding site (Fig. 4a). Intriguingly, NADH binding to FdsB differs from that to the homologous diaphorase unit of TtRC I (PDB-ID 3iam [https://doi.org/10.2210/pdb3IAM/ pdb]). Instead of stacking underneath the isoalloxazine ring of FMN, the nicotinamide moiety of NADH sits in front of the FMN binding pocket in hydrogen bond distance to the backbone of FdsB (Fig. 4b, c, Supplementary Movie 3). In this position it blocks access to the binding pocket and also prevents FMN from disengaging from the enzyme, but it is too far away (11.6 Å) to allow for productive electron transfer between NADH and FMN. In contrast, the coordination of the adenosine diphosphate (ADP) moiety of NADH resembles that of complex I. Conserved residues Glu 259 and Lys 157 form hydrogen bonds to the ribose oxygens of NADH and the alcohol groups of FMN contact the second phosphate group of ADP (Fig. 4b). Furthermore, the adenine stacks against Phe 152 of FdsB in analogy to Phe 70 of Nqo1 in TtRC I. Upon NADH binding to complex I, a hydrogen bond between Lys 202 and Glu 184 is broken and Lys 202 forms a bond to NADH 41  residues of NADH and may contribute to the non-productive positioning of the nicotinamide moiety. We have already shown that NADH can be used as an electron donor for the reduction of CO 2 to formate 6 . It is hence expected that NADH can bind the diaphorase unit of RcFDH productively to deliver electrons to the electron transfer chain via FMN. At 3.2 Å resolution we are unable to distinguish NADH from NAD + , but considering the estimated 100-fold excess of NADH over NAD + at the time of grid preparation and the fact that none of the residues in proximity of the nicotinamide nitrogen atom are suitable to stabilise a positive charge of NAD + , we have likely trapped NADH bound to the substrate inhibited enzyme in our cryo-EM structure.
Reduction of the RcFDH cofactors. When subtracting the map of the NADH reduced enzyme from that of the as isolated enzyme, the difference map essentially shows densities at all cofactors of the electron transfer chain (Fig. 4d). We conclude that the density in these regions of the as isolated EM structure are stronger than in the map of the NADH bound enzyme. This effect can either arise if the cofactors in the NADH bound enzyme move or if they show a different scattering behaviour due to charge. High-resolution crystallographic characterisation of the high-potential iron-sulfur protein in different redox states, indicate a small (up to 0.03 Å) contraction of the oxidised [4Fe-4S] cluster 42 with regard to the reduced cluster. These movements are too small to be visualised in our 3D reconstructions and would also not explain the observed densities around molybdenum, phosphates or FMN. EM maps reflect the charge of atoms 43 , as scattering of electrons by atoms in electron microscopy produces coulomb potential maps. In particular at resolution ranges between 5 and 10 Å, atomic scattering amplitudes are usually weaker the more negatively charged atoms are 44 . Hence, the difference map between the unsharpened EM maps of the as isolated and the NADH reduced states of FDH qualitatively visualises negative charges on electron accepting atoms. When the maps are B-factor sharpened to weigh down low resolution frequencies and boost high-resolution frequencies, the signal in the difference maps disappears, which is in agreement with the observation that scattering amplitudes differ very little at highresolution ranges (Supplementary Fig. 8). As observed in our EPR studies, NADH treatment generates Mo V states in a fraction of the sample, requiring electrons to travel along the entire electron transport pathway from FMN to Mo. The difference map provides a snapshot of all electron positions as an average charge change over all complexes in the EM analysis. Weak difference densities at the [2Fe-2S] clusters could arise due to fast electron transfer or reduction of these clusters in both states, though EPR spectroscopic characterization of the as-isolated state showed no reduction of Fe-S clusters. However, it cannot be ruled out that the [2Fe-2S] clusters may be partially reoxidised after NADH treatment, e.g. because they could be more prone to autooxidation during grid preparation. The difference density at [4Fe-4S] cluster A4 indicates that this cluster is likely redox active upon NADH reduction.

Discussion
We report the cryo-EM structure of the molybdoenzyme RcFDH.
It reveals an unexpected subunit composition as a dimer of FdsABGD heterotetramers. The arrangement of Fe-S clusters resembles that of complex I, supporting the idea that complex I and NAD + -dependent FDH evolved from the same ancestor 45 . Intriguingly, RcFDH can be loaded with electrons from the FMN binding site in the presence of NADH. The lack of structural changes at the bis-MGD pterin and dithiolenes indicates that redox changes of Mo VI to Mo V by NADH appear to principally involve the Mo metal ion. Our cryo-EM analysis of two different redox states of RcFDH shows that NADH reduction leads to charging of the cofactors in the absence of the second substrate at the bis-MGD. Since the experimentally obtained 3D density reflects the coulomb potential, our study proves that cryo-EM can indeed serve as a powerful tool to visualise charges on the cofactors of redox proteins, either by direct comparison of distinct redox states or by comparison of electron density maps with coulomb potential maps.

Methods
Cloning, protein expression and purification. The fdsGBACD operon was amplified using primers listed in Supplementary Table 6. The resulting fragment was cloned downstream of the nifH promotor with an N-terminal His 6 -tag before fdsG into vector pBBR1-MCS2 46,47 creating the plasmid designated pTHfds36. Protein expression was performed under anaerobic (photoheterotrophic) conditions in RCV medium 48  UV-visible and electron paramagnetic resonance spectroscopies. UV-visible spectra obtained for RcFDH were obtained either on a Shimadzu 1280 spectrophotometer housed in an anaerobic Coy chamber (Grass Lake, MI) (O 2 < 10 ppm) or aerobically on a Shimadzu 2600 spectrophotometer. Aerobically purified enzyme was brought into the anaerobic chamber and was made anaerobic via PD-10 buffer exchange columns (GE Healthcare) into degassed 100 mM Tris-HCl, 10 mM NaN 3 , pH 9.0 and was concentrated anaerobically to~500 µM using a centrifuge (1-15PK, Sigma, Germany) at 14,000×g. Enzyme used for aerobic experiments was prepared similarly using aerobic buffer. In all, 2-5 µM FDH was treated either with sodium formate (5 mM final concentration) or NADH (40 µM or 20 mM final concentration). Following this, the above sample was treated with sodium dithionite (2 mM final concentration). EPR samples were prepared aerobically in an ice bath using FDH purified as above, but afterward was desalted with a PD-10 column into in 100 mM Tris-HCl, 10 mM NaN 3 , pH 9.0. Typical EPR sample preparation methods involved addition of 20 µl of the above buffer or either freshly-prepared 100 mM NADH or sodium dithionite to 180 µl of as-isolated RcFDH residing in a quartz EPR capillary (3.9 mm O.D.), followed by brief mixing and relatively immediate freezing (10-15 s) in a liquid N 2 -cooled ethanol bath before final freezing in liquid N 2 . The final concentration of NADH or dithionite was 10 mM. CW X-band EPR spectra were obtained using a laboratory-built spectrometer (microwave bridge, ER041MR, Bruker; lock-in amplifier, SR810, Stanford Research Systems; microwave counter, 53181 A, Agilent Technologies) equipped with a Bruker SHQ resonator. An ESR 910 helium flow cryostat with an ITC503 temperature controller (Oxford Instruments) was used for temperature control. A Cu(II)/EDTA standard was used as a reference for spin quantitation of FDH samples 49 . Spin quantitation (by double integration) was performed using the utility 'spincounting' (https://github.com/lcts/spincounting) in Matlab (Mathworks). Magnetic field calibrations were applied through measuring a reference N@C 60 sample (g = 2.00204) at ambient temperature to compensate field offsets between Hall probe and sample position 50,51 . Parameters for EPR data acquisition at 12 K were: modulation amplitude, 5 G; microwave power, 4.0 mW; microwave frequency, 9.38 GHz. Parameters for EPR data acquisition at 80 K were identical, except that the modulation amplitude was 2 G.
Electron microscopy data acquisition. In all, 3.5 µl of as isolated RcFDH (100 µg/ ml) was either directly applied to freshly glow-discharged Quantifoil R2/4 300mesh holey carbon grids with 2 nm carbon support films or after 5 min incubation with 2 mM NADH. The protein solution was incubated on the grid for 45 s at 5°C and 85% humidity before blotting for 2 s and plunge freezing in liquid ethane using a FEI Vitrobot.
Cryo-EM images for the initial model were collected under low-dose conditions on a FEI Spirit microscope operated at 120 kV equipped with a 4kx4k F416 CMOS camera (TVIPS). We used CTFFIND4 52 for estimation of the contrast transfer function parameters, and Relion 1.4 53 or Imagic 54 for all subsequent steps. 1014 manually selected particles were subjected to 2D classification in Relion 1.4 in order to obtain references for template-based particle picking. Semi-automated particle selection, z-score sorting and 2D classification in Relion resulted in a FDH dataset of 116783 particles. Particle images were normalised, band-pass filtered between 200 and 10 Å, and classified using multivariate statistical analysis in IMAGIC. The class averages were used to generate an initial 3D reconstruction by angular reconstruction imposing two-fold symmetry. This reconstruction was submitted to 3D auto-refinement in Relion 1.4 resulting in a map with a final resolution of 14.13 Å.
High-resolution cryo-EM images of as isolated RcFDH were collected on a FEI Tecnai G2 Polara microscope operated at 300 kV equipped with a Gatan K2 Summit direct electron detector. 4623 micrographs were recorded in superresolution mode at a pixel size of 0.628 Å using LEGINON 55 . The defocus range was set from −0.6 to −3.2 µm. Each micrograph was dose-fractionated to 50 frames with a total exposure time of 10 s and a total dose of 64 e − /Å 2 . The first 25 frames were used for image processing.
High-resolution cryo-EM images of NADH incubated RcFDH sample were collected on a FEI Tecnai G2 Polara microscope operated at 300 kV equipped with a Gatan K2 Summit direct electron detector and on a FEI Titan Krios microscope operated at 300 kV equipped with a Gatan K2 Summit direct electron detector. On the Polara microscope, 4082 micrographs were recorded in super-resolution mode at a pixel size of 0.628 Å using LEGINON. The defocus range was set from −0.6 to −3.5 µm. Each micrograph was dose-fractionated to 50 frames with a total exposure time of 10 s and a total dose of 64 e − /Å 2 . On the Krios microscope, 3984 micrographs were recorded in counting mode at a pixel size 1.1 Å. The defocus range was set from −0.6 to −3.2 µm. Each micrograph was dose-fractionated to 50 frames with a total exposure time of 8 s and a total dose of 40 e − /Å 2 . The first 25 frames were used for image processing.
Cryo-EM image analysis. Image processing and 3D reconstruction was performed using RELION-3.0 56 . Movie frame alignment and dose-weighting was performed with MotionCor2 57 and contrast transfer functions were determined using CTFFIND4. All refinements used gold standard Fourier shell correlation (FSC) calculations and reported resolutions are based on the FSC = 0.143 criterion of mask corrected FSC curves. All maps were masked and sharpened using automatically determined negative B-factors. Supplementary Figures 2 and 6 as well as Supplementary Table 5 show all steps of image processing of each of the datasets. Channel cavities were detected using the CAVER 3.0.1 58 plugin for PyMOL (The PyMOL Molecular Graphics System, Version 2.0 Schrödinger, LLC.) with following settings: probe radius, 0.97; shell radius, 3.0; shell depth, 8.0; frame weighting coefficient, 1.0; frame clustering threshold, 1.0. Micrographs of as isolated RcFDH sample showing strong astigmatism, over focus, very low defocus, broken ice or ice contaminations were discarded, resulting in 4347 micrographs for further processing steps. 2D classification of 1237 manually picked particles generated templates for semi-automated particle selection. The dataset of 1117847 particles was subjected to z-score sorting and several iterative 2D classifications to remove bad particles, resulting in a final dataset of 799023 RcFDH particles. 3D classification with exhaustive angular searches were performed to further clean the dataset resulting in one good class with 366558 particles which was used for subsequent 3D auto-refine. The initial model was filtered to 40 Å and used as a reference for 3D auto-refinement of the entire dataset resulting in a RcFDH reconstruction at 3.43 Å resolution without applied symmetry and 3.30 Å resolution with applied C2 symmetry, respectively. CTF per particle Refinement of the C2 symmetry map resulted in 3.26 Å resolution and was used for model building.
All frames were used for image processing. Micrographs of NADH incubated RcFDH sample showing strong astigmatism, over focus, very low defocus, broken ice or ice contaminations were discarded, resulting in 4082 micrographs of the FEI Polara dataset and 3216 micrographs of the FEI Titan Krios dataset for further processing steps. For semi-automated particle selection, the templates of the oxidised dataset were adjusted to the corresponding boxsize and pixelsize. The dataset of 1215305 (FEI Polara dataset) and 744234 (FEI Titan Krios dataset) particles was subjected to z-score sorting and several iterative 2D classifications to remove bad particles, resulting in a final dataset of 669964 RcFDH particles from FEI Polara dataset and 370195 RcFDH particles from FEI Titan Krios dataset. At this point the two datasets were merged resulting in a total pf 1037182 particles with a customised pixel size of 1.07 Å. To ensure the two datasets were merged properly an additional round of 2D classification was done resulting in a final dataset of 669976 RcFDH particles. 3D classification with exhaustive angular searches were performed to further clean the dataset resulting in one good class with 199229 particles which was used for subsequent 3D auto-refine. The initial model was filtered to 40 Å and used as a reference for 3D auto-refinement of the entire dataset resulting in a RcFDH reconstruction at 3.57 Å resolution without applied symmetry and 3.37 Å resolution with applied C2 symmetry, respectively. NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-15614-0 ARTICLE NATURE COMMUNICATIONS | (2020) 11:1912 | https://doi.org/10.1038/s41467-020-15614-0 | www.nature.com/naturecommunications CTF per particle Refinement of the C2 symmetry map resulted in 3.24 Å resolution and was used for model building.
Generation of difference maps. In contrast to scattering of X-rays by atoms, electron scattering amplitudes heavily depend on the charge of the atom at low resolution ranges 43,[59][60][61] . When two experimental EM maps representing different redox states of the same complex are subtracted from each other, the difference map will visualise charge differences. This effect is strongest at resolution ranges between 5 and 10 Å, so that difference maps will indicate the charge better when produced from unsharpened maps than from sharpened maps. Difference maps were produced by subtracting refined maps prior to post-processing or after Bfactor sharpening as indicated using the vop subtract command in Chimera with the option -minRMS to normalise the data. Only data obtained from the Tecnai G2 Polara microscope were used for the generation of difference maps.
Molecular modelling. Both maps show clear side chain density for almost all residues allowing for model building by homology model-guided chain tracing. Model building of RcFDH as isolated complex was carried out in COOT 62 into the different refined density maps. Crystal structures of sequence homologous structures were used to guide model building for FdsA, FdsG and C-terminal part of FdsB. FdsD and the N-terminal part of FdsB were built de novo into density maps based on defined densities of bulky residues. For cofactor addition, unmodeled blobs were identified and filled with the respective ligand. The models were refined using phenix.real_space_refine implemented in PHENIX using additional geometry restraints for the fitted ligands. The data collection and model statistics are summarised in Supplementary Table 5. Our atomic model covers residues 7-955, 1-493, 2-149 and 2-70 in FdsA, FdsB, FdsG and FdsD, respectively. We thus obtained a structural model for almost the entire RcFDH complex.
Alignments. Clustal Omega 63 was used for sequence alignment. Jalview 2.11.0 64 was used for visualizations.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The EM maps of the as isolated structure and the structure in the presence of NADH are deposited under accession codes EMD-10496 and EMD-10495, respectively. Atomic