Structure of the bacterial flagellar rotor MS-ring: a minimum inventory/maximum diversity system

The bacterial flagellum is a complex, self-assembling, nanomachine that confers motility on the cell. Despite great variation across species, all flagella are ultimately constructed from a helical propellor attached to a motor embedded in the inner membrane. The motor consists of a series of stator units surrounding a central rotor made up of two ring complexes, the MS-ring and the C-ring. Despite many studies, high resolution structural information is still completely lacking for the MS-ring of the rotor, and proposed mismatches in stoichiometry between the two rings have long provided a source of confusion for the field. We here present structures of the Salmonella MS-ring, revealing an unprecedented level of inter- and intra-chain symmetry variation that provides a structural explanation for the ability of the MS-ring to function as a complex and elegant interface between the two main functions of the flagellum, protein secretion and rotation.

The flagellum is the organelle responsible for the swimming motility of a huge variety of bacterial 27 species, many of which are of clinical relevance, and the driving force behind this swimming ability has 28 fascinated researchers since it was first observed in the 17 th century 1 . Flagella are highly complex, 29 being formed from more than 25 different proteins assembled into a series of circularly symmetric 30 and helical assemblies 2-4 . Electron cryotomographic (cryo-ET) studies have demonstrated that 31 flagellar structures are hugely variable across species, depending on whether the flagella are to be 32 located freely in the extracellular environment, encased in an outer membrane sheath, or entirely 33 within the periplasm 5 . At the core of every flagellum, however, is a highly conserved inner-membrane 34 motor that is attached to a drive-shaft, which ultimately culminates in the flagellum (Fig. 1a). The 1 motor itself consists of a rotor complex surrounded by stator proteins that are proposed to generate 2 torque. Rotation is rapid (up to 1,700 Hz in some species 6 ), utilises ion flow through the membrane 3 (usually a proton-motive or sodium-motive force 7 ), and can respond dynamically to chemotactic 4 signals 8 . The stators transmit the torque to a cytoplasmic complex known as the C-ring that consists 5 of three proteins (FliG, FliM, FliN 9,10 ). The N-terminal domain of FliG interacts with the extreme C-6 terminus of FliF 11,12 , which in turn forms the MS-ring, a large predominantly periplasmic structure that 7 is tethered to the inner membrane via N-and C-terminal transmembrane helices 13,14 . In addition to 8 interfacing to the C-ring to form the rotor, the MS-ring is one of the first flagellar structures to 9 assemble 15 , and houses the type III secretion system (T3SS) of the flagellum that is responsible for the 10 secretion and assembly of the helical components forming the drive-shaft and propellor 16 . The  ring therefore sits at the heart of the flagellum, both structurally and functionally. Despite this, little 12 is known about its structure. Salmonella enterica serovar Typhimurium (S. Typhimurium) FliF consists 13 of 560 amino acids with predicted transmembrane helices close to the N-and C-termini (Fig. 1b). 14 Sequence analysis of residues 50-460, which lie in the periplasm, predicted that FliF consists of a series 15 of ring building motifs (RBMs) that have previously been observed in periplasmic ring forming proteins 16 of related secretion systems 17 . Most notably, the prediction for RBM3 was that it is formed from two 17 disparate stretches of sequence, with a long insertion between two of the predicted β-strands. Early 18 estimates of FliF stoichiometry from purified S. Typhimurium flagella suggested approximately 27 19 copies per flagellum 18 , and low resolution electron cryomicroscopy (cryo-EM) studies produced 20 reconstructions with 24-,25-and 26-fold rotational symmetries applied 19,20 . Similar analyses of the S. 21 Typhimurium C-ring also showed variable stoichiometry, but centred on a 34-fold symmetry 20-22 . 22 Models of flagellar rotation have so far needed to account for these proposed mismatches in 23 symmetry, with disagreements over the exact location of the mismatch and the functional 24 implications (summarised in 23 . We here present near-atomic resolution structures of the MS-ring 25 from S. Typhimurium that resolve these issues, revealing a conservation of stoichiometry between 26 MS-and C-rings and unusual internal symmetry mismatches that account for the multiple functions 27 of the rotor. 28 29

FliF forms rings of mixed internal symmetry 30
In order to better understand how a single protein could perform multiple different roles, we over-31 expressed and purified the S. Typhimurium MS ring (FliF), using modifications of previously published 32 protocols 13 , and collected single-particle cryoEM data from Triton X-100, DDM and amphipol A8-35 33 solubilised protein (Extended Data Fig. 1). Analysis of near top-down views after 2D classification (Fig.  34 1d) revealed a periodicity at the extremity of the largest ring consistent with > 30 subunits, rather than 1 the expected 25-27. Ab-initio reconstructions and 3D-classifications were performed using a variety 2 of imposed symmetries, but initially only a C33 reconstruction produced an interpretable map. 3 Further refinement of this model led to a 2.6 Å volume (Fig 1c,  33 copies of residues 231-438 of FliF, corresponding to the RBM3/β-collar, were built de novo (Fig.  5 1e). However, all other densities within the C33 volume could not be interpreted as protein, suggesting 6 either high levels of disorder or different symmetries (Extended Data Fig. 2). We therefore used the 7 C33 particle set to perform a reconstruction in C1 which, at low resolution, revealed a periodicity 8 underneath the C33 ring consistent with C21 symmetry. Masked refinement of this region with C21 9 symmetry imposed led to a 2.9 Å reconstruction (Fig. 1c, e, f; Extended Data Table 1) in which residues 10 125-222, corresponding to RBM2, could be built. Further refinement of these particles imposing the 11 common C3 symmetry of the two main rings revealed density consistent with a further nine copies of 12 RBM2 decorating the outside of the 21-fold symmetric RBM2inner ring (Fig. 1g). Underneath each copy 13 of RBM2outer is a smaller density with secondary structures features consistent with a homology model 14 for RBM1 (Fig. 1g). The overall visible structure therefore contains twenty one copies of FliF 15 contributing one RBM3 to the 33-fold ring and one RBM2inner to the 21-fold ring, nine copies of FliF 16 contributing one RBM3 to the 33-fold ring and one RBM2outer, and three copies of FliF only contributing 17 one RBM3 to the 33-fold ring (Fig. 1h). In addition we observe density consistent with nine copies of 18 RBM1 and, although we assume they pair with the nine copies of RBM2outer, we cannot see linking 19 residues to confirm this connectivity. We see no clear protein density that accounts for the remaining 20 three RBM2s, the remaining twenty four RBM1s, or any of the transmembrane helices and cytoplasmic 21 portions, but we found no clear evidence of proteolytic fragments in gels (Extended Data Fig. 1) or by 22 proteomic analyses (data not shown). We did however observe further diffuse densities including a 23 ring of material close to the observed C-terminal residues, which is consistent with detergent micelle, 24 and a column of weak density in the centre of the structure underneath the RBM2inner ring (Extended 25 Data Fig. 3).  FliF that make up one third of the complex reveal the small changes in relative orientations of the 14 RBM2 and RBM3 domains between different copies required to build the full object. 15 16 17 1

FliF monomer structures 2
The enormous complexity of the 33-fold MS-ring means that there are a variety of monomer 3 structures, with each of the 11 chains in the nominal asymmetric unit being unique in terms of relative 4 domain orientation (Fig. 1h). Each chain, however, is made up of equivalent domains that match the 5 predicted structural arrangement well (Fig. 1b). The density that corresponds to RBM1 fits a homology 6 model based on domain 1 of the type III secretion system (T3SS) injectisome protein SctJ, with a βαββα 7 topology. RBM2 and RBM3 are both canonical RBM domains with an αββαβ topology. Despite The total surface area buried in the FliF ring is enormous (217000 Å 2 ), totalling 36 % of the available 31 monomer surface. All of the interaction surfaces observed in the assembly are highly conserved, with 32 areas of greatest variation occurring in surface loops and disordered regions (Extended Data Fig. 8). 33 Analysis of the electrostatic surface potential of the monomers reveals that the interaction surfaces 1 are mostly hydrophobic, but patches of complementary charge are observed (Extended Data Fig. 9). 2 3 The complex can be broken down into four main structural assemblies: the 33-fold symmetric β-collar, 4 the 33-fold symmetric RBM3 ring, the 21-fold symmetric RBM2inner ring and the decorating 5 RBM2outer/RBM1 domains. The β-collar accounts for 77000 Å 2 of the buried area and consists of 66 6 vertical β-strands (shear number of 0) linked to 66 β-strands angled ~60° from the horizontal. In 7 addition to the standard β-sheet hydrogen bond network, this sub-structure is stabilised by numerous 8 sidechain mediated hydrogen bonds and two potential salt bridges (His281-Asp369 and Glu280-9 Arg370). In addition, we observed a density connecting Arg373 and Lys275 from neighbouring 10 subunits, consistent with a glutaraldehyde cross-link formed during the final purification stage before 11 imaging (Extended Data Fig. 10). Arg248 and Arg154-Glu139 inter-subunit salt bridges respectively (Fig. 3c). With the exception of the 19 linkers between them, there is virtually no contact between the RBM3 and RBM2inner rings (Extended 20 Data Fig. 12). This is likely a reflection of the fact that the symmetry mismatch between the rings both 21 prevents a consistent interaction surface and leads to a significantly smaller diameter for the RBM2inner 22 ring (70 Å) compared with the RBM3 ring (140 Å) or the β-collar (100 Å). Contacts between the two 23 mis-matched rings are instead via the RBM2outer subunits. The C-terminal loop of one RBM2outer subunit 24 tucks between two RBM3 subunits in the ring above, while the second α-helix of the αββαβ motif 25 bridges two RBM2inner subunits rings (Extended Data Fig. 12). Additional contacts are observed 26 between the C-terminus of an RBM1 domain and two of the RBM2inner subunits. However, the lower 27 resolution of these portions of the structure suggests these are not tight contacts. 28

29
The mechanism by which such a complex arrangement of subunits and mixed symmetries could be 30 built from a single protein chain is intriguing. It appears that the interaction surfaces of the RBM3 and 31 RBM2 domains are primed to build rings of significantly different stoichiometry, and hence there is a 32 need to build in flexibility and the extreme symmetry breaking innovation of the RBM2outer 33 conformations. The majority of the structure is built from units containing two copies of the RBM2inner 34 conformation and one copy of the RBM2outer conformation, presumably driven by limitations of the 1 conformations the linkers can take preventing more than two consecutive RBM2inner conformations. 2 These blocks could either be visualised as an RBM2inner/RBM2outer/RBM2inner arrangement, in which the 3 RBM2outer subunit bridges the two RBM2inner, or as an RBM2inner/RBM2inner/RBM2outer arrangement, in 4 which case the RBM2outer provides the bridge to the next unit. This pattern is observed for three such 5 units (contributing 9 RBM3s to the 33-fold ring and 6 RBM2inners to the 21-fold ring), at which point 6 the pattern is broken by a subunit pair containing one copy of an RBM2inner conformation and one 7 copy for which there is no visible RBM2 density. This completes one third of the structure and this 8 pattern of packing is then repeated twice more. 9 10 FliF combines elements of sporulation and secretion system structures 11 The closest structural homologue of the 33-fold RBM3 domain assembly is the SpoIIIAG protein from 12 the sporulation system of B. subtilis 25 , which forms a 30-fold symmetric structure utilising a very 13 similar interaction surface to the RBM3 domains of FliF ( Fig. 2a and Extended Data Fig. 13). Strikingly, 14 SpoIIIAG also contains a β-insertion that forms a 60-strand β-collar with a shear number of 0, and both 15 proteins also share a feature of a short triangular insertion at the point the strands change direction 16 to the vertical (Fig. 2a). Such vertical β -strands are highly unusual, but have also been observed in the 17 outer membrane secretin structures of T3SS and type II secretion systems (T2SS) 24,26 . Despite the 18 strong similarities, there are significant differences between FliF and SpoIIIAG, most notably in the 19 angle of the RBM domain to the β-collar (Fig. 2a, Extended Data Fig. 13). 20

21
The RBM2 domain is most closely related at both sequence and structural levels to the RBM2 domain 22 of the SctJ family from the virulence T3SS injectisomes and again this homology extends to the ring 23 structures formed (Fig. 2b). The inner-membrane proximal portion of an injectisome basal body is 24 formed by two different proteins, SctD and SctJ. The inner SctJ ring has been shown to house the 25 export gate structure of the secretion system in its central cavity 27,28 and forms a 24-fold symmetric 26 ring from RBM1 and RBM2 domains 24 . The interaction interface between neighbouring domains in 27 the SctJ RBM2 ring is closely related to the FliF RBM2inner packing interaction (Extended Data Fig. 14), 28 although the copy number difference does lead to a small difference in the size of the cavity. SpoIIIAG (PDB-5wc3). A view from the outer-membrane side is shown above and from the side below, 4 with a cartoon representation of a single, extracted, monomer also shown. A small beta-insertion 5 structure is indicated (arrow on monomer structures). b, The virulence T3SS basal body is constructed 6 from two protein chains in the MS-ring equivalent region, which both form 24-mer rings consisting of 7 multiple RBM domains. Central sections of the Salmonella SPI-1 injectisome basal body (PDB-5tcr, LH 8 panel) and the FliF ring (RH panel) show the striking similarity in overall shape despite fundamental 9 differences in the chains and domain types used. They also show the 21-fold RBM2inner domains are 10 very similarly arranged to the PrgK/SctJ-RBM2 24-fold ring, whilst the FliF-RBM1 and PrgK/SctJ-RBM1 11 domains are very differently arranged with respect to these. 12 13

FliF exists in multiple stoichiometries 14
More detailed analyses of 3D classifications of the FliF particles imposing C3 symmetry revealed a 15 subset of particles in which the C33 features were subtly broken. Further classification of these 16 particles in C1 only allowing local angular sampling revealed that they corresponded to a C34 17 symmetric MS-ring. Refinement of this volume with C34 symmetry imposed led to a 2.8 Å structure of 18 the RBM3/β-collar region of this form of the MS-ring (Fig. 3a). Initial attempts to reconstruct the 19 RBM2inner region of these particles with C21 symmetry were unsuccessful and so the C34 20 reconstruction was used as a reference in a C1 reconstruction that revealed 22-fold symmetry. Masked 21 refinement of the RBM2inner region with C22 symmetry imposed led to a 3.1 Å structure of this portion, 22 while refinement of the whole volume with the common C2 symmetry led to a 3.3 Å reconstruction 1 ( Fig. 3a and Extended Data Fig. 15). Surrounding the 22-fold symmetric RBM2inner ring we observed 2 ten densities that correspond to the RBM2outer/RBM1 domain pairs observed in the 33mer structure, 3 but again no density was observed for the final two copies of RBM2 or for twenty four copies of RBM1. 4

5
The MS ring is therefore capable of assembling into rings of differing stoichiometries. Analysis of the 6 interfaces buried in the FliF 34mer revealed that only very subtle changes are needed to build the 7 alternate stoichiometry (Fig. 3b, c). The interfaces used in the 33-fold RBM3/β-collar ring are identical 8 to those in the 34-fold RBM3, maintaining all of the bonding interactions, including the salt bridges 9 (Fig. 3c). A similar pattern is observed for the RBM2inner rings. Although the changes are subtle, when 10 propagated around the number of copies in the ring, they do make a difference to the diameter of 11 each ring, with a 2.3 Å (2 %) increase seen in the RBM3/β-collar ring and a 6.6 Å (9 %) increase observed 12 for the RBM2inner ring. The most significant difference between the two structures exists in the 13 RBM2outer/RBM1 domain pairs, where an extra copy is observed. However, the mode of packing of the 14 RBM2outer/RBM1 against the RBM2inner ring and the basic 2:1 (RBM2inner:RBM2outer) building block is 15 conserved. The larger rings permit five copies of the trimer building block to assemble before the 16 pattern is broken by the minority 1:1 (RBM2inner:RBM3only) block. It is worth noting that the 17 symmetries of the two main rings always leave twelve copies of the monomer which don't contribute 18 to RBM2inner and twenty four copies of RBM1 for which we see no density, although the significance 19 of these observations is currently unclear. 20 21 Once we had observed two different assemblies in our sample, we attempted to assess whether other 22 symmetries were also present at lower levels. To achieve this we performed supervised 3D 23 classifications in C1 using reference models generated to reflect RBM3 symmetries from C32 to C36 24 (Extended Data Fig. 16). This analysis confirmed that the majority of the particles partitioned into the 25 C33 and C34 classes (40% and 23% respectively), with 7% and 9% ending up in the C32 and C35 classes 26 respectively. The remaining 20% went into the C36 class, but reconstructions of these particles were 27 very low resolution and clearly artefactual. elsewhere. The colour scheme mimics that of Figure 1c. b, Comparison of the C33/C34 and C21/C22 5 regions by overlaying the complete rings using a single chain reveals the subtle differences in the sizes 6 of the respective ring-like assemblies built. c, Despite assembling to form rings of different 7 symmetries, the specific interactions from which they are built are entirely conserved, including salt 8 bridges, in both the C33 (cyan and blue)/C34 (grey) rings (left hand panel) and the C21 (cyan and 9 blue)/C22 (grey)rings (right hand panel). 10 11 The MS-ring as structural adaptor 1 The structural heterogeneity observed in this study may seem surprising for a core component of such 2 a fundamental cellular structure, but agrees with earlier demonstrations of stoichiometric 3 heterogeneity for the S. Typhimurium C-ring 20-22 . The C-ring is a large cytoplasmic structure that 4 assembles on to the MS-ring via a mechanism in which the N-terminal domain of the first C-ring 5 protein, FliG, folds around two helices at the C-terminus of FliF 11,12 . The other domains of FliG then 6 recruit the other C-ring components, FliM and FliN 29-32 , as well as providing the interaction surface for 7 the stator complexes that generate torque 33, 34 . The MS-ring/C-ring junction is therefore critical for despite this region of the volume being averaged with 25-fold symmetry (Fig. 4a). Although we do not 21 observe the C-terminal residues of FliF in our structure, the positioning of the RBM3 domains on the 22 outside of the ring mean they are correctly placed to reach down to the FliG ring underneath the 23 membrane. This observation was confirmed by placing the FliF structure into a subtomogram average 24 of in situ flagella from Plesiomonas shigelloides (Fig. 4b) 35 . Interestingly this placement also provides 25 further insights into other roles the symmetric complexity of the MS-ring may play in acting as a single 26 chain structural adaptor molecule at the centre of the system (Fig. 4c). The RBM3 domains of the 27 structure, and hence the cytoplasmic C-termini, have the 33/34-fold symmetry required to assemble 28 the C-ring. The RBM2inner domains, on the other hand, form the 21/22-fold symmetric ring that is seen 29 to house the export gate in the homologous injectisome structures 24,27,36 . The highly conserved 30 dimensions of the export gate 37 compared to the large diversity in C-ring size between bacterial 31 species drives the requirement for symmetry mismatch between the different domains of FliF. The 32 subtle differences in size between the central pore of the RBM2inner 21/22mers and the equivalent 33 24mer SctJ injectisome ring suggests there is some flexibility in the details of how the export gate is 34 accommodated, perhaps related to the differences within the inner membrane region located below 1 this ring seen when comparing cryo-ET of flagella and injectisomes 38  showing the good match in overall shape and links to the 34-fold symmetric C-ring. The 34-mer FliF 16 was built in the map shown in Figure 3a and extended to the C-terminus using a continuous helix of 17 the correct length, ending in a homology model based on the crystal structure of residues 523-559 of 18 Helicobacter pylori FliF (PDB: 5wuj). b, The FliF model (coloured as in (a)) is shown placed in a P. 19 shigelloides tomographic volume (EMD-10057) and a model for the export gate complex (blue) (PDB-20 6r69) is then docked within FliF. The panel on the right is an update of the cartoon from Figure 1a, 21 using this colour scheme. c, Exploded diagram of FliF coloured to emphasise the roles the different 22 symmetries play in adapting between components within the flagellar assembly. 23 Conclusion 24 This study has provided, for the first time, a near-atomic resolution view of the MS-ring of the bacterial 1 flagellar rotor. The structures reveal unexpected symmetries and an unprecedented level of structural 2 heterogeneity for a homo-oligomeric assembly. The symmetry mismatches within the structure 3 demonstrate how the MS-ring is able to bridge multiple different structural and functional units within 4 the flagellar basal body utilising a single protein chain (Fig. 4c). The explicit linking of the MS-ring 5 stoichiometry to that of the C-ring introduces new questions of how rotors of different sizes in 6 different species of bacteria can be reconciled with this model, especially given the constraints that 7 the need to house the T3SS in the centre of the structure places on the system. Will FliF provide yet 8 more surprises or will other adaptor proteins play a role? 9 10 Acknowledgements 11 We thank E. Johnson and A. Costin of the Central Oxford Structural Microscopy and Imaging Centre 12 for assistance with data collection. H. Elmlund (Monash) is thanked for access to SIMPLE code ahead 13 of release. We thank Morgan Beeby (Imperial College) for access prior to publication, to the P. Frozen cell pellet was resuspended in 40 ml of lysis buffer (50 mM Tris pH 8, 50 mM NaCl, 5 mM EDTA) 1 and lysed by 3 passes through an Emulsiflex C5 homogeniser (Avestin) at 10,000 psi. After 2 centrifugation at 20,000 x g for 20 min to remove cell debris, cell membranes were collected by 3 ultracentrifugation at 186,000 x g for 1 hour. Collected membranes were dissolved in 40 ml of alkaline 4 buffer (50 mM CAPS pH 11, 5 mM EDTA, 50 mM NaCl, 1 % (w/v) DDM) at 4 °C for 1 hour. Undissolved 5 material was removed by centrifugation at 20,000 x g for 20 minutes. Solublised FliF was then pelleted 6 by ultracentrifugation at 143,000 x g for 1 hour. Pelleted FliF was resuspended in 2 ml of resuspension 7 buffer (25 mM HEPES pH 8, 50 mM NaCl, 0.1 % (w/v) DDM). FliF ring assemblies were then separated Micrographs were initially processed in real time using the SIMPLE pipeline 44 , using SIMPLE-unblur 32 for motion correction, SIMPLE-CTFFIND for CTF estimation and SIMPLE-picker for particle picking. 33 Following initial 2D classification in SIMPLE to remove poor quality particles, all subsequent processing 34 was carried out in in RELION-3.0 45 . Particles were re-extracted using a 432 x 432 pixel box from 1 micrographs that had been re-processed using the MotionCor2 46 implementation in RELION-3.0, with 2 CTF estimation by CTFFIND4 47 . 3 4 Initial processing of the Triton-X100 extracted particles produced 2D classes with close to top down 5 views that allowed preliminary counting of the subunits around the perimeter of the object, although 6 the lack of purely top down views prevented unambiguous assignment. 27435 particles were selected 7 after classification and used to generate ab initio initial models with C33 and C34 symmetry. 3D 8 classification was carried out with C33 and C34 symmetries applied and the C33 job produced a class 9 containing 15634 particles that refined to 3.8 Å using gold standard refinement. Reclassification of the 10 original particles produced a 19520 particle set that led to a 3.1 Å map following Bayesian polishing 48 11 and per-particle CTF refinement. This allowed de novo model building of the RBM3/β-collar domains 12 (residues 231-438) but all other regions of the map remained untraceable. Attempts at reconstructing 13 with lower symmetry were hindered by the low particle number. 14 15 A larger, DDM-extracted, dataset was collected that contained 188007 particles after 2D classification. 16 3D classification applying C33 symmetry resulted in one good class with 106745 particles which were 17 then used in a C1 symmetry refinement. This produced density in the ring below the C33 ring with a 18 clear periodicity that could be counted as C21. Refinement of this particle set with the common 19 symmetry of C3 applied produced a 3.3 Å map, following Bayesian polishing and CTF refinement, that 20 revealed an RBM fold in the 21-fold symmetric ring. However, the quality of this portion of the map 21 was not sufficiently detailed to allow de novo model building. As the proportion of particles that 22 produced a sub-3.5 Å map were similar between the two different detergent extractions, and the 23 maps produced were indistinguishable, we created a combined dataset containing the post-2D 24 classification particles from the Triton-X100 and DDM extractions and a small dataset from a DDM 25 extracted sample that had been exchanged into amphipol A8-35. This dataset, containing 273493 26 particles was subjected to 3D classification applying C3 symmetry, using the DDM-only model low pass 27 filtered as a reference. After two rounds of 3D classification, two good classes were produced, 28 containing 126285 and 59163 particles. The first of these classes refined to a pure C33 object in the 29 RBM3 region, but the second class produced a map with ~11.3 subunits per "asymmetric unit" in the 30 C3 symmetry. We therefore re-refined this class applying C34 symmetry, which produced a 3.3 Å gold 31 standard map. Refinement of the "C34" particles in C1 produced periodicity in the ring below the 32 RBM3 ring consistent with C22 symmetry. 33 34 Due to the increased complexity of the sample, we collected a large A8-35 exchanged dataset and 1 created a composite Triton-X100/DDM/A8-35 dataset containing 449142 particles after 2D 2 classification. These particles were then subjected to a supervised 3D classification in C1, using C33 3 and C34 maps as references, producing classes with 308536 and 140606 particles respectively. The 4 C33 class was subjected to a further round of classification, producing a good class with 175233 5 particles that was refined in C3 to an overall resolution of 2.9 Å following Bayesian polishing and CTF 6 refinement. Further focused classification and refinement of the C33 particles with a mask around the 7 RBM3/β-collar region, and with C33 symmetry applied, produced a 2.6 Å map from 77849 particles. 8 Further focused classification and refinement of the C33 particles with a mask around the RBM2inner 9 region, and with C21 symmetry applied, produced a 2.9 Å map from 84797 particles. Attempts to 10 improve the resolution of the RBM2outer/RBM1 region through particle subtraction, multi-body 11 refinement and local averaging were unsuccessful. Initial refinements of the entire object produced 12 maps with nine strong copies of the RBM2outer/RBM1 pair and weaker density in the gaps between 13 copies 3 and 4, 6 and 7 and 9 and 1. This weaker density was consistent with being a superposition of 14 two copies of the RBM2outer/RBM1 density. However, the spacing of these domains was such that 15 these gaps could not accommodate a full RBM2outer/RBM1 pair without structural rearrangement, and 16 we reasoned that the density observed could be produced by rotational misalignment of a subset of 17 the particles producing "ghost" density from the strong domains. In order to test this, we masked 18 around the nine strong domain pairs and used this mask in a focused refinement, with the logic that 19 if extra copies were genuinely ordered they would appear in the final, unmasked, map. This was not 20 found to be the case. The C34 class was refined with C2 symmetry applied and produced a 3.3 Å map 21 after Bayesian polishing and CTF refinement. Further focused refinement of the C34 particles with a 22 mask around the RBM3/β-collar region, and with C34 symmetry applied, produced a 2.8 Å map. 23 Further focused classification and refinement of the C34 particles with a mask around the RBM2inner 24 region, and with C22 symmetry applied, produced a 3.1 Å map from 87107 particles. Similar analysis 25 of the RBM2outer/RBM1 region was applied as in the C33 refinements, with similar results, but in this 26 case ten copies of the domain pair could be placed. All processing statistics are summarised in 27 Extended Data Tables 1 and 2.  28 29 Model building and refinement 30 A monomer model for the RBM3 and the β-collar (residues 231-438) was built manually in Coot 49 31 using the 2.6 Å map with C33 symmetry applied, assembled into a 33-fold model, and refined using 32 phenix.real_space_refine 50 . A monomer model for RBM2 (residues 125-222) was built manually in 33 Coot using the 2.9 Å map with C21 symmetry applied, assembled into a 21mer of the RBM2inner region 34 and refined using phenix.real_space_refine. The whole 33mer was assembled from these two 1 structures in the 2.9 Å map with C3 symmetry applied. The two main rings were joined by manually 2 building the linkers in Coot. Nine copies of the high resolution RBM2 domain were placed manually in 3 the RBMouter domain densities of a 4 Å lowpass filtered version of the C3 map, and rigid body refined. 4 Nine copies of a RaptorX generated homology model of RBM1 (residues 50-106) were manually 5 positioned in the density underneath the RBMouter domains and rigid body refined. The completed 6 33mer was refined with phenix.real_space_refine, using the higher resolution C33 and C21 structures 7 as reference models. The RBM3/β-collar monomer built in the C33 map was used to assemble a 34-8 fold model in the 2.8 Å map with C34 symmetry applied, and was refined using 9 phenix.real_space_refine. A 22-fold RBM2inner model was assembled in the 3.1 Å map with C22 10 symmetry applied, using the RBM2 monomer built in the C21 map, and was refined using 11 phenix.real_space_refine. The whole 34mer was assembled from these two structures in the 3.3 Å 12 map with C2 symmetry applied. The two main rings were joined by manually building the linkers in 13 Coot. Ten copies of the RBMouter/RBM1 domain pairs from the 33mer model were placed manually in 14 the appropriate densities of a 4 Å lowpass filtered version of the C2 map, and were rigid body refined. 15 The completed 34mer was refined with phenix.real_space_refine, using the higher resolution C34 and 16 C22 structures as reference models. All models were validated using Molprobity 51 . All refinement and Leeuwenhoek