Structural cycle of the Thermus thermophilus PilF ATPase: the powering of type IVa pilus assembly

Type IV pili are responsible for a diverse range of functions, including twitching motility and cell adhesion. Assembly of the pilus fiber is driven by a cytoplasmic ATPase: it interacts with an inner membrane complex of biogenesis proteins which, in turn, bind to nascent pilin subunits and mediate fiber assembly. Here we report the structural characterization of the PilF TFP assembly ATPase from Thermus thermophilus. The crystal structure of a recombinant C-terminal fragment of PilF revealed bound, unhydrolysed ATP, although the full length complex was enzymatically active. 3D reconstructions were carried out by single particle cryoelectron microscopy for full length apoprotein PilF and in complex with AMPPNP. The structure forms an hourglass-like shape, with the ATPase domains in one half and the N1 domains in the second half which, we propose, interact with the other pilus biogenesis components. Molecular models for both forms were generated: binding of AMPPNP causes an upward shift of the N1 domains towards the ATPase domains of ~8 Å. We advocate a model in which ATP hydrolysis is linked to displacement of the N1 domains which is associated with lifting pilin subunits out of the inner membrane, and provide the activation energy needed to form the pilus fiber.


Results
Crystal structure of the C-terminal portion of the TtPilF assembly ATPase. Crystallization trials on the complete TtPilF hexamer were unsuccessful; however, limited proteolysis gave rise to a fragment of approximate molecular mass 44 kDa which was purified and yielded crystals which diffracted to 2.44 Å (Supplementary  Table S1). Mass spectrometry was used to confirm that this fragment corresponded to the C-terminal half of the protein, encompassing the N2, CTD and Zn-binding domains (designated TtPilF c ). The structure was solved using phases obtained from SeMet-labelled protein (Supplementary Table S1), which was necessary to assist in a molecular replacement solution of the structure.
TtPilF c forms a flattened hexamer, with 2-fold rotational non-crystallographic symmetry relating the two trimers (Fig. 1A). TtPilF c is identical in sequence but in a different crystal form to a structure published previously for a similar TtPilF fragment derived from a different Thermus thermophilus strain 18 . An alignment of both structures gave an r.m.s.d. of 0.354 Å; an overlay of TtPilF c with PilB from Geobacter metallireducens, which also has 2-fold rotational symmetry within the hexamer, gave an r.m.s.d. of 1.564 Å ( Supplementary Fig. S1A,B). Although the two forms of TtPilF c crystallised in different space groups, their crystal packing was very similar ( Supplementary Fig. S1C,D). Inspection of all 6 binding sites revealed electron density consistent with bound nucleotide; however, there was no evidence for ATP hydrolysis in any site ( Supplementary Fig. S2). ATPase assays were carried out on the TtPilF c fragment which showed that it was not enzymatically active, confirming that the fragment is catalytically inert in solution, as well as in the crystalline state. Superposition of all chains showed that the conformation of ATP/Mg 2+ in the 6 binding sites is similar ( Supplementary Fig. S3). This observation shows that the ATP conformation is unaffected by differences in contacts between adjacent subunits caused by flattening of the hexamer (Fig. 1A). Recognition of bound ATP was similar for all chains-details of interacting residues are shown in Fig. 1B. Thr651, Gly652, Ser653, Lys655 and Ser656 form part of the Walker A motif involved in binding the γ and β phosphates. These residues, as expected, are highly conserved within the PilB family 17 .
The principal difference between the PilFc structure reported here and that presented by Mancl et al. 18 , is that the latter contained phosphothiophosphoric acid-adenylate ester (ATPγS), an ATP analog with S substituted for O in the γ phosphate. ATPγS is a substrate for most protein kinases, although the activity levels are probably considerably lower than for the natural ATP substrate 24 . The complex with ATPγS also gave no indication of hydrolysis of the terminal phosphate, in keeping with our observations. We systematically compared the conformations of the two ligands in both structures for each equivalent binding site (Fig. 1C). Mg 2+ occupies the position of the thiophosphate moiety in chains A, B, E and F; for the remaining two chains, the conformation of both nucleotides is very similar, with Mg 2+ ions bound to the ATPγS. Furthermore, where density for Mg 2+ ions was present in the ATPγS complex, we noticed that it was weaker than the six copies in the ATP complex. We conclude that the variation in conformation of bound ATPγS between different sites is attributable to the weaker electronegativity of S compared with O, reducing the polarity of the S = P bond, compared with O = P and changing the local charge distribution.
One possible explanation for the failure of the TtPilF c fragment to hydrolyse ATP is the conformation of the linker polypeptide between the CTD and N2 domains. A comparison of TtPilF c chain A with equivalent subunits from Geobacter metallireducens PilB in ADP-and AMPPNP-bound forms showed significant deviations in the linker region ( Supplementary Fig. S4). Small changes in the orientation of catalytic residues within the active site can have a dramatic effect on catalytic efficiency.

Cryoelectron microscopy structures of TtPilF in the apoprotein and AMPPNP-bound states.
Cryoelectron microscopy was used in conjunction with single particle averaging to generate 3D reconstructions of the complete TtPilF hexamer, as an apoprotein and in complex with AMPPNP (sample data are shown in Supplementary Fig. S5). We have previously reported the structures of TtPilF in the two states previously, but at lower resolution 25 . In our previous work we clearly identified the complex as a hexamer, but at lower resolution there was no evidence of a C2 symmetry deviation: we therefore refined the structures of both forms of TtPilF with C6 symmetry. However, it became apparent during the processing of the superior direct detector data that a subtle C2 symmetry was apparent in the complex top-view (see exemplar class averages in Supplementary  Fig. S5C). Furthermore, a C2 symmetry designation would be consistent with the crystal structures of the ATPase domains discussed previously. This was obvious during the processing stages and application of C6 symmetry-produced maps, with good correspondence between class averages and structure back projections only for side views. Estimation of the resolution of the TtPilF map (AMPPNP-bound form) by FSC correlation gave a value of 8 Å at the conservative value of 0.5 ( Supplementary Fig. S6A). Calculation of the map resolution by ResMap 26 showed an even distribution of resolution through the height of the complex, and suggested a higher overall resolution value of 4-5 Å ( Supplementary Fig. S6B,D). Visual inspection of the map suggested that the more conservative value from the FSC correlation is likely to be closer to the correct value. The same calculations carried out on the TtPilF apoprotein map gave a similar resolution estimate from the FSC value, although the ResMap estimate was slightly worse, at 5-6 Å.
The electron density maps obtained were similar in their general features to those determined previously 25 , but with considerably more detail; an example of the map quality for the AMPPNP-bound state is shown in Fig. 2. The structure divided into two approximately equal masses although, at this improved resolution, it was readily apparent that they are different. There was also much more detail around the central region, which appears more complex than the simple 'stem' structure which was observed previously 25 . Of particular note is the fact that the part of the structure associated with the ATPase domains is more distorted from C6 symmetry than the N1 domain section (Fig. 2B). A comparison of the two density maps revealed that they are similar in overall shape and dimensions, but the apoprotein structure is shorter and wider than the AMPPNP structure, indicative of conformational change between the two states (discussed further below).

Comparison of the structures of the ATPase domains.
To construct a molecular model for the TtPilF hexamer, we examined the fitting of the TtPilF c crystal structure into the AMPPNP cryoelectron density map. Two alternative positions were possible which were evaluated independently by rigid body fitting ( Supplementary  Fig. S7). The correctly docked form was readily identified by a superior correlation coefficient (0.83 versus 0.70), which reflected a much better fit of the crystal structure into the top half of the density map ( Supplementary  Fig. S7). The fit of the docked model was optimized by refinement in ROSETTA 27 , giving the model TtPilF AMPPNP . The TtPilF AMPPNP structure was then used as a starting point for refinement against the apoprotein map, and the final model was designated TtPilF apo ; structures of both forms are shown in Fig. 3. Electron density was complete for almost all of both models in this region of the density map.
Significant differences were apparent between the TtPilF AMPPNP model and the TtPilF c crystal structure. Superposition of the CTD and N2 domains gave an r.m.s. displacement of 3.47 Å for Cα atoms; when viewed from the top, TtPilF c is wider by about 6 Å along the long axis and shorter by 7 Å along the shorter axis, leading to a flatter hexagonal shape (Fig. 4A). Given that the hexamer adopts C2 symmetry, three chains were compared (F, C and B) which lie on one side of the hexamer. Superposition of the CTD domains showed that the relative orientations of the N2 domains were largely preserved but with translational shifts, depending on the chain ( Supplementary Fig. S8). To quantify these movements, we determined the centers of mass for each domain and compared shifts in distances from domains on opposite sides of each hexamer ( Table 1). The CTDs for chains A and B, at the widest point of the hexamer, are further apart in the crystal structure, but the reverse is the case for chains C and D, reflecting the flatter structure of the TtPilF c hexamer. The N2 domains, however, were all consistently further apart in TtPilF c compared with TtPilF AMPPNP . Differences in the relative positions of CTD and N2 domains on each chain, which presumably also alter the conformation of the linker which contributes to ATP binding, could explain the activation of ATPase activity in the intact TtPilF hexamer, compared with the enzymatically inert TtPilF c .
A similar approach was used to compare the ATPase domains in TtPilF AMPPNP and TtPilF apo . The apoprotein structure showed a consistent contraction towards the center of the complex for all domains-readily apparent from a superposition of both structures (Fig. 4B) and also an analysis of the displacement of the centers of mass (Table 1 and Fig. 4C). Displacement of the N2 domains appeared to be generally greater than the CTDs: this observation is significant because it is this domain which is linked to the N1 domains in the N-terminal half of TtPilF. It therefore appears that binding of AMPPNP to TtPilF apo triggers an expansion of the hexamer, outwards from the center.  ADP-and AMPPNP-bound forms of PilB G, using PHENIX 29 , gave an r.m.s. deviation of 1.4 Å. In other words, the two conformations of PilB G are more similar to each other than they are to the ATPase domains of TtPilF in either state. We conclude, therefore, that the transition from TtPilF apo to TtPilF AMPPNP is better described by an outward shift in center of mass, particularly of the N2 domains, rather than changes in the relative orientation of adjacent CTD/N2 units. Such a difference may be attributable to the fact that electron density for the N1 domains is detected in our structures and could influence structural transitions observed within the ATPase domains. In addition, it is important to note that our cryoelectron microscopy-generated maps are not at a sufficient resolution to determine whether all three non-identical ATP binding sites are occupied by AMPPNP, or whether there are differences in catalytic competence and/or ligand binding between non-identical sites.
Modelling of N1 domains. No crystal structures are currently available for the N1 domains from TtPilF but structures of the equivalent domains from type II secretion system ATPases are available 30,31 . Each TtPilF    50 . Chains are labelled in Fig. 4A.
chain forms three N1 domains, which we will term the first, second and third, numbered from the N-to the C-terminus 25 . To dock the N1 domains into the AMPPNP density map, we first employed the segmentation facility SEGGER 32 , as implemented within Chimera 33 , to identify potential domain boundaries. The results were indicative of 12 separate domains in the lower half of the PilF hexamer i.e. two domains per TtPilF chain. It was also noteworthy that the domains formed two separate rings-an upper and a lower. We reasoned that the upper ring most likely corresponded to the third N1 domain in the TtPilF sequence, which lies adjacent to the ATPase domains. The lower ring would therefore correspond to the second N1 domain. Clearly, this implies that the density for the first N1 domain is not well resolved, possibly because it is weaker and/or unfolded to some degree. This might be because it plays a role in interaction with TtPilM (see below).
To identify the orientation of each N1 domain, models for the second and third N1 domains were generated using I-TASSER 34 , based on the structure of the MshE N-terminal domain from Vibrio cholerae ( Supplementary  Fig. S9) 31 . For each N1 domain, a comprehensive 6D search was conducted against each electron density map for the 6 separate, segmented domains. For each ring in the complete TtPilF structure, a consistent orientation of each domain was identified within the 8 highest scores for each search. The complete domains were then subject to refinement of fitting by ROSETTA 27 . This rigid body model was then transferred to the apoprotein electron density map and re-refined in ROSETTA to generate the model for the apoprotein form. Both models gave excellent agreement with experimentally determined electron density for each form (Fig. 5).
The MshE N-terminal domain from Vibrio cholerae, which forms the basis for the models of the N1 domains in TtPilF, consists of two subdomains: a four α-helix bundle and a 3-stranded β-sheet packed against 3 additional helices 31 . Collectively, the N1 domains stack into a cylinder, with external dimensions of about 130 Å wide by 55 Å in height. The model of the TtPilF cylinder is formed from two rings, each containing 6 N1 domains; the N1 domains are orientated such that the four α-helix subdomains lie on the outside of the cylinder (Fig. 6A). The α/β subdomains line the core, which forms a channel approximately 40 Å in internal diameter. In the model, the docked N1 domains in the upper and lower rings are proposed to be related by quasi-two fold symmetry of the four helix subdomains, with the symmetry axis perpendicular to the main C2 symmetry axis of the complex. The major site of interaction between adjacent N1 domains in the upper and lower rings is the fourth helix in the four α-helix sub-domain and the linker peptide which connects the two domains (Fig. 6B).

Movement of N1 domains in response to structural changes in the ATPase domains. In order
to examine the motion of the N1 domains relative to the ATPase domains in TtPilF AMPPNP and TtPilF apo , the ATPase domains from TtPilF AMPPNP were superimposed onto their equivalents from TtPilF apo 28 . Calculations of centers of mass for each domain showed that the TtPilF AMPPNP N1 domains from both ring 1 and ring 2 are displaced downwards by distances of 4-9 Å (Fig. 7). This is accompanied by rotation outwards (mean value 14 o for ring 2 domains), which leads to a wider aperture at the base of the structure. Combined, these movements lead to changes in the relative positions of the N1 domains in ring 2 of 10-13 Å which occur on the transition from TtPilF AMPPNP to TtPilF apo and therefore, by implication, after hydrolysis of ATP. Conversely, this means that the N1 and ATPase domains are drawn closer together in the apoprotein state. It seems likely that the movement of the N2 domains plays a critical role in this structural transition although, at this resolution, we were unable to directly trace the path of the polypeptide chain from each N2 domain to its N1 counterpart. The precise way in which the relative movement of the CTD and N2 domains is transmitted to the N1 domains in ring 1 therefore remains unclear. However, we note that the central mass of density which lies on the symmetry axis, which we termed the 'stem' in our previous work 25 , is not accounted for in our models and could be attributed to the missing section of the model between the N1 and N2 domains. This would confirm our previous hypothesis-that this central part of the TtPilF hexamer plays a critical role in linking structural transitions in the ATPase domains which occur on binding and hydrolysis of ATP with displacement of the N1 domains in the N-terminal region of the molecule.

Discussion
ATPases play a critical role in powering both TFP biogenesis and type II secretion, but the precise manner by which ATP hydrolysis is coupled to secretion and fiber assembly is still unclear. They are members of the much larger family of AAA+ ATPases which are involved, not just in secretion, but also a range of other cellular processes, such as chaperone-related and protein folding functions 14 . Phylogenetic analysis shows that the PilB/GspE subgroup, which is associated with type IVa pilus biogenesis and type II secretion, forms a distinct clade separate from the BfpF and the PilT/PilU clades 17 . Structural studies on the T2SS ATPase GspE have been complicated by difficulties in obtaining an assembled hexamer, requiring engineering of the assembly into a hexameric state 19 . By contrast, T. thermophilus PilF can be isolated as an intact hexamer and is somewhat unusual, in that it harbours 3 predicted N1 domains in the N-terminal half of the molecule. However, we found density for only 12 such domains, leading us to conclude that electron density for the first N1 domains from each of the 6 chains in the complex is weak in  35 . This is consistent with our structural model, which shows contacts between adjacent GSPII/N1 domains in the hexamer-disruption of these contacts could destabilise assembly. The GSPII/N1 domains were found to be not required for ATPase activity but, interestingly, mutation of the GSPII domains had effects on transformation efficiency and twitching motility 35 .
Addition of pilin subunits from the inner membrane into a growing or retracting pilus fiber would require sequential insertion or removal from the fiber base in a clockwise or counter-clockwise manner. Models for ATPases which drive assembly or disassembly therefore need to incorporate an explanation of how this phenomenon is linked to ATP hydrolysis. Nivaskumar et al. analyzed pilus fibers formed from the type II secretion pilin PulG, finding experimental evidence for variation in the twist angle, which was confirmed by computation, leading to a three stage pilus assembly model 7 . The first assembly stage is accompanied by a displacement of 10.4 Å, which removes the pilin from the inner membrane in a spooling-type mechanism. Interestingly, this displacement value is similar to the shift of the N1 domains relative to the ATPase domains which we observe on binding of AMPPNP. We find that TtPilF c is wider and flatter than its equivalent density in the complete hexamer in the AMPPNP-bound state, indicating that the presence of the N1 domains does affect the relative orientations of the ATPase domains. McCallum et al. compared the structures of the Geobacter metallireducens PilB ATPase in AMPPNP and ADP-bound states, leading to a model in which pairs of protomers hydrolyse ATP in a sequential clockwise direction 17 . However, neither of the PilB structures had clear electron density for the N1 domains, which could influence the conformational states of the ATP-binding domains within the complex. The authors also reinterpreted structures of the PilT retraction ATPase, originally published by Satyshur et al. 20 , according to the same model which suggested that PilT employs a similar mechanism, but operating in the reverse direction, thus disassembling a right-handed pilus helix. Despite the lower resolutions obtained, an advantage of the single particle approach is that it allows a consideration of the complete TtPilF structure and thus avoids potential concerns about crystal packing artefacts.
Our observations on TtPilF structures provide little indication of a rotary mechanism. Binding of AMPPNP is accompanied by an outward expansion of the CTD and N2 domains which is linked to a downward and outwards displacement of N1 domains on binding of AMPPNP (Fig. 7). A key question is, how might this be related to displacement of a pilin subunit out of the inner membrane? A displacement-type mechanism would require anchoring of TtPilF to the inner membrane, or components within it. Current evidence, from other TFP biogenesis systems, suggests that PilF interacts with the inner membrane protein PilC and the PilMNO complex, which could provide the anchoring necessary. Formation of such a complex could provide the environment for engagement with a membrane-bound pilin, reducing the activation energy necessary to guide it out and into the nascent pilus fiber. Possibly the first N-terminal N1 domains, which we do not observe in our structures, could play a key role here, by engagement with inner membrane components. Although there is no complete structure of a PilB-type ATPase with defined N1 domains yet reported, a complex of GspE bound to GspL-an inner membrane component of the T2SS-has been published 36 . Here the CTD and N2 domains adopt an unusual orientation such that, when compared with their counterparts from TtPilF, the N2 domains lie on different sides of the CTD (Supplementary Fig. S10). The N1 domain, as well as GspL, lies close to the CTD in GspE. This is a radically different arrangement from TtPilF, where there is relatively little contact between the CTD and N2 domains on the one hand and the N1 domains on the other in the TtPilF complex-they are apparently held together by the stem structure discussed earlier. However, GspE in the Vibrio complex does not form a hexamer in the crystalline state, which could contribute to differences in the relative locations of the constituent ATPase domains.
Recent results obtained by cryoelectron tomography have provided valuable insights into the TFP biogenesis machine in vivo 21-23 . Of particular relevance here is the study by Gold et al., which determined the structures of the TFP machinery in open and closed states 21 . Although the PilQ secretin and other periplasmic components are well resolved in the tomograms, the parts of the assembly platform in the inner membrane and cytoplasm are less clear. This is also the case for the tomograms from Myxococcus 22 and Vibrio 23 and has been taken as an indication of a dynamic assembly platform at the cytoplasmic surface of the inner membrane 37 . Clearly there is a need for a dynamic interchange between PilB-type assembly ATPases and retractile PilT ATPases; contacts with the inner membrane components of the machine will need to be exchangeable. PilB (or TtPilF) is thought to operate through binding to the PilC integral membrane protein 38 and PilM, which forms a complex with PilN and PilO 10 . Further work will be required to determine how structural changes in TtPilF are communicated to PilC and PilMNO, and what role these other components play in mediating type IVa pilus fiber assembly.

Methods
Expression and purification of PilF and PilF c . The full length Thermus thermophilus HB8 pilF gene was amplified from genomic DNA using PCR and cloned into the pET15b vector as described previously 25 . The final construct contained a 6×-His tag followed by thrombin cleavage site at the N-terminus of PilF. The plasmid was transformed into T7 express cells (New England Biolabs) and few colonies were used to inoculate a 100 mL of 2×-YT media containing ampicillin (100 μg/ml) and grown at 37 °C for 3 hours. The starter culture was diluted into 4 L of fresh 2×-YT media and the cells were allowed to grow at 37 °C until the OD at 600 nm reached 0.6-0.8. At this stage, the culture was cooled to 16 °C and IPTG was added to a final concentration of 0.1 mM and allowed to grow for 16 hours. The cells were harvested by centrifugation at 6000 g for 10 min, resuspended in lysis buffer (25 mM HEPES pH 7.5, 150 mM NaCl, 5 mM MgCl 2 and 5% glycerol (v/v)) and lysed by sonication. The debris was removed by centrifugation at 30,000 g for 30 min, the supernatant was passed through a 0.45 μm filter and followed by a 0.2 μm filter. The filtered solution was injected into a 5 ml HisTrap column (GE Healthcare) using a peristaltic pump. The column was washed with lysis buffer containing 20 to 100 mM imidazole pH 7.5 by step gradients of 3 column volumes each. Increasing the imidazole concentration to 500 mM eluted the PilF protein. The protein was incubated at 50 °C for 15 min and then centrifuged at 18000 g for 15 min and the supernatant was collected.
For cryoelectron microscopy studies, the full length PilF from the previous step was concentrated and injected into a Hiload Superdex 200 (16/60) column (GE Healthcare) equilibrated with lysis buffer. The pure fractions from the elution were pooled and concentrated to 1 mg/ml. To form the PilF-AMPPNP complex, AMPPNP was added to a final concentration of 1 mM.
For crystallization studies, the full length PilF from the metal chelate affinity and heating steps was concentrated to ~4 mL, and then ATP was added to a final concentration of 2 mM. To this, trypsin was added (PilF:trypsin ratio of 1:100 (w/w)) and incubated at 37 °C for 30 min. Powdered complete protease inhibitor tablet (Roche) was added to the sample to stop the trypsin digestion and then injected into a Hiload Superdex 200 (16/60) column equilibrated with lysis buffer supplemented with 2 mM ATP. The pure fractions from the elution were pooled and concentrated to 16 mg/ml. Analysis by SDS-PAGE and mass spectrometry revealed almost all of the N-terminal GSPII domains were digested by the trypsin treatment.
For selenomethionine (SeMet) labeling, the cells were grown in SeMet media (Molecular Dimensions Ltd) as per the manufacture's guidelines. The cells were lysed and the protein was purified as explained for the native protein except that the lysis buffer was supplemented with 1 mM DTT throughout.
ATPase assay. The ATPase activity was measured using a phosphate assay kit (Abcam), following the manufacturer's guidelines. Samples (200 µl) containing TtPilF (30 µg) were mixed with ATP (55 µM) and incubated at 55 o C for 15 min. The phosphate reagent (30 µl) from the kit was then added to the samples and incubated for 30 min at room temperature, protected from light. Optical density at 650 nm was measured for samples using a microplate reader (Biotek). Experiments were carried out in triplicate.
Crystallization and crystal structure determination. PilF c was screened against eight commercial crystallization screens (Molecular Dimension and Microlytic) using the MRC 2-well crystallization plates and Mosquito robot (TTP labtech). The crystallization plates were incubated at 20 °C and plate-like crystals appeared within one week. PilF c crystals belonged to C2 space group and grew in 0.2 M sodium bromide, 0.1 M Bis-Tris propane pH 6.5, 20% (w/v) PEG 3350. The presence of a translation non-crystallography symmetry, combined with diffraction anisotropy in one direction complicated the identification of an unambiguous molecular replacement (MR). During this time, structures of the T2SS ATPases GspE and EpsE and the T4P retraction ATPase PilT were the only available search models for MR. The optimal quality dataset was obtained from a crystal treated with K 2 PtCl 4 , as part of attempts to extract phase information.
Our efforts to obtain experimental phasing information by collecting data on native crystals soaked with heavy atoms were only partially effective: SeMet incorporated protein was therefore purified and crystallized. Although the SeMet crystals diffracted only to 4.0 Å resolution, an unambiguous solution was found by combining SAD and SCIEnTIFIC REPORTS | (2018) 8:14022 | DOI:10.1038/s41598-018-32218-3 MR with PilT coordinates (PDB 2GSZ, three chains) as the search model in Phaser 39 . This weak solution was then extended to the higher resolution dataset (crystal treated with K 2 PtCN 4 ), which also belonged to the same crystal form as the SeMet dataset. The models were built using the Autobuild wizard 40 as incorporated in Phenix 29 . The model was completed by iterative rounds of manual building in Coot 41 and refinement using phenix.refine 42 . The model was optimized using PDB_REDO 43 and validated using molprobity tools 44 . We did not observe any conformational changes between the low resolution SeMet dataset and the high resolution dataset. The data and refinement statistics are presented in Supplementary Table 1 and a Ramachandran plot is given in Supplementary  Fig. S11. Cryoelectron microscopy. 3 μl of samples at a concentration of ~200 μg/ml where absorbed to freshly glow discharged gold 200 mesh Quantifoil R2/2 grids using an FEI Vitrobot. After a brief incubation at 90% humidity, grids were blotted continuously for 4-6 seconds and plunge frozen in liquid ethane. Data were acquired on an FEI Titan Krios at the electron Bio-Imaging Centre at Diamond Light Source. Movies were acquired in counting mode on a Gatan GIF Quantum LS at 1.38 Å/pixel and a 20 eV energy filter was used. The dose rate for movie acquisition was 5e/pixel/s and a total dose of ~40 e/Å2 was accumulated over 32 movie frames. Individual frames were corrected for drift and motion then summed with dose weighting using MotionCorr 45 . Approximately 40,000 clear particles were semi-automatically picked using E2Boxer and initial reference free class averages were generated in EMAN 46 and independently cross-validated using RELION 47 . Bad particles and contaminated classes were removed and subsequent class averaging clearly identified a hexameric particle with a distinctive C2 oligomeric arrangement and therefore a low resolution start model was generated with this symmetry applied. 6 rounds of iterative refinement were performed to produce 3D volumes.
Structure refinement into electron density. Refinement was carried out using the ROSETTA relax protocol 27 , with constraints against the electron density map, but without B-factor refinement. 100 separate simulations with either the ATPase domains or combined N1 domains were carried out against electron density maps of either the apoprotein or AMPPNP-bound forms. The lowest energy models were selected in each case. Figures were generated using Chimera 33 and CCP4MG 48 .

Data Availability Statement
Coordinates and structure factors for the TtPilF c crystal structure were deposited with accession code 5OIU. Electron density maps and fitted models were deposited with the following accession numbers: AMPPNP-bound form, EMD-4194, PDB 6F8L; apoprotein form, EMD-3882, PDB 6EJF. Other data are available in the article and associated Supplementary Information Files.