We first characterized IMPDH2 filaments assembled in vitro by addition of ATP (Anthony et al., 2017). The affinity of IMPDH2 for ATP has not been directly measured but we found ATP concentrations as low as 1 μM sufficient to induce assembly (Figure 2A). We prepared cryo-EM grids of ATP-assembled filaments and found that in the absence of other ligands, IMPDH2 filaments are extremely flexible (Figure 2—figure supplement 1A). Two-dimensional class averages confirmed our prior observation from negative stain that the filaments are composed of conformationally heterogeneous octamers stacked head-to-head, resulting in flexible filaments with variable rise and radius of curvature (Figure 2B). Because these deviations from ideal helical symmetry severely limited attempts at image processing by traditional iterative helical real-space reconstruction, we attempted to produce more structurally homogeneous filaments by addition of IMP or NAD⁺, which stabilize the flexible active site loops (Sintchak et al., 1996). As previously reported, substrates did not have a direct effect on IMPDH2 filament assembly, and filament assembly did not directly affect enzymatic activity (Figure 2C–D) (Anthony et al., 2017). While the turnover rate we observe here (~0.05 per second) is somewhat low, these reactions were performed near room temperature (24°C); we found that increasing reaction temperature to 37°C resulted in increasing Kcat to ~0.1 per second, roughly half what was observed in another recently published study of human IMPDH2 (Figure 2—figure supplement 2) (Fernández-Justel et al., 2019). Unfortunately, addition of IMP and NAD⁺, either alone or in combination, did not significantly reduce filament flexibility (Figure 2—figure supplement 1B–D). Two-dimensional class averages of helical segments again exhibited varying degrees of curvature and showed no correlation of structural states between neighboring IMPDH2 octamers, preventing successful three-dimensional processing with conventional helical approaches (Figure 2E–G).

From these 2D class averages we observed that filament flexibility was due to variations in the conformation of filament segments, but that the interface between segments did not vary, and as a consequence this region was better resolved (Figure 3A). Focused refinement of the filament assembly interface region alone provided a valuable foothold in resolving high resolution structures of all regions of the filaments. We developed a workflow for a single-particle style approach to reconstruction of inherently heterogeneous helical protomers, combining density subtraction, symmetry expansion, focused classification and focused refinement to isolate, classify, and reconstruct discrete sections of filaments for structural characterization (Figure 2—figure supplement 1E–G). This approach provided multiple structures for two different focused regions of IMPDH2 filaments: full octameric segments, and the paired catalytic domain tetramers that constitute the filament assembly interface (Figure 3B).

To define the interactions that drive assembly of catalytically active IMPDH2 filaments, we obtained cryo-EM structures of IMPDH2 in three liganded states: 1) ATP/IMP/NAD⁺, 2) ATP/NAD⁺, and 3) ATP/IMP. The three datasets were qualitatively similar. However, because the structures refined to higher resolution when both substrates were present, we focus our analysis here on the ATP/IMP/NAD⁺ dataset. The complete image processing workflow applied to this dataset (Figure 3—figure supplement 1A) was also applied to the two single-substrate datasets. From the ATP/IMP/NAD⁺ data we resolved many different structural classes, three best-resolved of which were: the filament assembly interface, and two different reconstructions of octameric filament segments (Figure 3—source data 1). We resolved many conformations of filament segments, however for all segments the filament assembly interface was identical. Therefore the filament assembly interface structure is a dataset consensus structure, an average of every segment included in the dataset. We did not observe any conformational cooperativity in the compressed/extended conformational equilibrium between IMPDH2 octamers sharing a filament assembly interface.

We determined the structure of the ATP/IMP/NAD⁺ consensus filament assembly interface at 3.0 Å resolution (Figure 3C, Figure 3—figure supplement 1B–C, Figure 3—figure supplement 2A). This D4 symmetric region is composed of two symmetrically opposed catalytic domain tetramers. The interface between tetramers is formed by the 12 amino-terminal residues of eight protomers, which each extend from the core of the molecule to bind one catalytic domain on the opposite tetramer, in a shallow surface groove formed by a short helix (476-485), two beta strands (51-63), and two short loops (355-360, 379-380) (Figure 3D). A key tyrosine/arginine interaction (Y12/R356’) anchors the attachment; mutation of either of these residues to alanine was previously shown to abolish assembly, both in vivo and in vitro (Anthony et al., 2017). Residues 1–7 make mostly hydrophobic interactions with the catalytic domain, and an embedded arginine (R480’) is positioned to hydrogen bond to the I6 backbone carbonyl. These interactions are reciprocal - a protomer extends its N-terminus to a partner and receives the N-terminus of the same partner, creating four pairs of symmetrical interactions at the interface. The overall surface area buried by a single interaction is 2,300 Ų, with a total buried area of 9,200 Ų for the multimeric interface.

The N-terminal tail of IMPDH, corresponding to residues 1–28 in IMPDH2, is the least conserved part of the protein, with large variation in length and sequence between phyla, as well as conformational variation among known structures (Figure 3—figure supplement 3A) (Buey et al., 2015a; Fernández-Justel et al., 2019; Kim et al., 2017; Labesse et al., 2013; Makowska-Grzyska et al., 2015; Osipiuk et al., 2014; Prosise and Luecke, 2003; Trapero et al., 2018). The residues involved in filament contacts, however, are conserved among chordates, consistent with the fact that IMPDH polymerization has only been reported in vertebrates (Figure 3—figure supplement 3B). In previous crystal structures of human and fungal IMPDH, residues 1–11 are unresolved and residues 12–28 are well resolved. In these structures, Asp16 anchors the tail in place through ionic interactions with Arg341/Lys349 of a neighboring protomer, while Val13 and Pro14 make hydrophobic contacts. In the filament structure, however, the Val13/Pro14 contacts are broken and the entire tail is rotated about 30° to position Tyr12 to contact Arg356 across the filament interface (Figure 3—figure supplement 3C).

In this consensus structure, the remainder of the catalytic domain is nearly identical to a structure of human hIMPDH2 bound to competitive inhibitors in a previously determined crystal structure (PDB 1nf7, backbone RMSD 0.641 Å for residues 18–107, 245–417, and 439–514) (Sintchak et al., 1996). The active site is well-resolved, except for one loop (residues 421–436). There is strong density in the both the IMP and NAD⁺ binding sites. We have modeled these ligands as IMP and NAD⁺, however because the enzyme filaments were actively turning over when flash frozen, we have likely captured a mixture of substrate-, intermediate-, and product-bound states. Attempts at focused classification of the active site to structurally isolate these states were unsuccessful.

We resolved multiple conformations of filament segments in the ATP/IMP/NAD⁺ dataset (Figure 3—figure supplement 1D). The most well-resolved was a D4-symmetric, fully extended octamer (3.3 Å, Figure 3E, Figure 3—figure supplement 1E–H, Figure 3—figure supplement 2B). The eight Bateman domains are symmetrically extended with respect to their catalytic domains, giving the octameric segment a helical rise of 118 Å; a symmetric helix of these segments would have a twist of 36°. There is strong density at both of the canonical ATP-binding sites within the Bateman domain (Figure 3F), consistent with previous crystal structures of ATP/ADP-bound fungal and bacterial IMPDH (Buey et al., 2017; Labesse et al., 2013).

We also resolved several bent filament segment structures, the best resolved of which reached 3.9 Å (Figure 3G, Figure 3—figure supplement 1I–L, Figure 3—figure supplement 2C). This bent octamer contains two identical tetramers. For each, three protomers are in extended conformations, while one protomer is in the compressed conformation. Due to the symmetry of the octamer, the lone compressed protomer of each tetramer forms a Bateman domain dimer with an extended protomer of the opposing tetramer (Figure 3H, Figure 3—figure supplement 3D). From each of the lower resolution ATP/IMP and ATP/NAD⁺ datasets, we also observed a single consensus filament assembly interface and many different filament segment classes, including fully extended and bent segments, as well as fully compressed segments (Figure 3I–J, Figure 3—source data 2, Figure 3—source data 3). To our knowledge, these are the first structure of IMPDH in the compressed conformation in the absence of guanine nucleotides. While the two canonical ATP sites are occupied, the third Bateman binding site (which forms only in the compressed state and is GTP/GDP-specific) remains unoccupied. This suggests that protomers within ATP-bound IMPDH filaments readily sample the compressed conformation, and that GTP binding to site 3 selectively stabilizes this state. Further, the ability of a compressed protomer to form a Bateman dimer with an extended partner demonstrates a lack of conformational cooperativity across the Bateman domain interface.

We and others previously reported that GTP can stabilize compressed IMPDH2 filaments or drive their disassembly, depending on what other ligands are present (Anthony et al., 2017; Buey et al. (2017); Duong-Ly et al., 2018). To understand how ligand status of the active site tunes the response to GTP, we systematically explored the effects of different ligand combinations on filament assembly. In the absence of substrate, GTP induces disassembly of filaments, but as little as 10 μM IMP inhibits disassembly (Figure 4A–B). Pre-treatment of filaments with IMP prevents disassembly by GTP, and addition of IMP promotes reassembly of filaments previously disassembled by GTP (Figure 4—figure supplement 1A–B).

To understand how IMP and GTP allosterically influence filament assembly and disassembly of ATP-bound IMPDH2, we acquired cryo-EM data of the enzyme in two ligand states: 1) ATP/GTP/IMP, and 2) ATP/GTP. To ensure morphological consistency, we sought to saturate the enzyme with GTP. For the ATP/GTP dataset, we used 2 mM GTP, which for both our initial negative stain experiments and cryo-EM preparations resulted in complete disassembly of filaments into free octamers (Figure 4—figure supplement 1C). However, under saturating IMP concentrations, 2 mM GTP resulted in filaments that were often bent (Figure 4—figure supplement 1D). This should not be possible if GTP were saturating all sites; our structures above suggest bent filaments must contain some extended protomers whose GTP-binding allosteric site 3 is disrupted. For the ATP/GTP/IMP cryo-EM dataset we therefore used a much higher GTP concentration (20 mM), which resulted in filaments with fully compressed segments (Figure 4—figure supplement 1E). We explore this apparent difference in GTP affinity in more detail below.

The compressed filament dataset resulted in three high-resolution reconstructions: the consensus filament assembly interface, a fully compressed octameric filament segment, and a fully compressed octameric end segment (Figure 4—figure supplement 2A, Figure 4—source data 1). The ATP/GTP/IMP filament assembly interface map (3.0 Å) is identical (backbone RMSD = 0.407) to the ATP/IMP/NAD⁺ filament interface (Figure 4—figure supplement 2B–C, Figure 4—figure supplement 3A). Classifying the most well-resolved filament segments as before, we obtained a 3.2 Å structure of a fully compressed filament segment (Figure 4C, Figure 4—figure supplement 2D–H, Figure 4—figure supplement 3B). The Bateman domains are symmetrically compressed, with ligand density at all three allosteric sites. The active site is well resolved, including the canonical interactions between opposing active site fingers that inhibit catalysis by preventing catalytic dynamics (Buey et al., 2017).

Compared to the uninhibited filament datasets, these GTP-saturated, +IMP filaments were shorter in length. As a result, in addition to filament segments, we identified many filament ends: octamers in which one tetramer does not have an assembled interface. The best-resolved structure of these filament ends (3.3 Å) is conformationally similar to the filament segments, being fully compressed, with well-resolved IMP-bound active sites in the inhibited conformation; however, without the filament assembly interface, the N-terminus is only partially resolved (Figure 4D, Figure 4—figure supplement 2I–L, Figure 4—figure supplement 3C).

From the free octamer dataset containing ATP/GTP, we used a similar symmetry-expansion and classification strategy to that used for the filament datasets (Figure 4—figure supplement 4A). This scheme confirmed that virtually all the particles were symmetric, fully compressed free octamers in which both potential filament assembly interfaces are unbound, with some poorly resolved minority classes of larger oligomers (Figure 4—figure supplement 4B–C). The best resolved free octamer class reached intermediate resolution (4.5 Å) (Figure 4E, Figure 4—figure supplement 4D–F, Figure 4—figure supplement 5, Figure 4—source data 2). This octamer is conformationally similar (backbone RMSD = 0.978 Å) to a recent higher-resolution crystal structure of IMPDH2 bound to GTP (PDB 6i0o) (Buey et al., 2017; Fernández-Justel et al., 2019). The Bateman domains are fully compressed with three occupied allosteric sites. Without IMP present, the active site loops of the free octamers are disordered, with partial density for the fingers extending across to the opposing tetramer.

Whether in a filament segment, filament end, or free octamer, the Bateman domains of the three ATP/GTP-bound IMPDH octamer types are fully compressed, with ligand density in all regulatory sites including the critical GTP/GDP-specific third site, which ‘staples’ IMPDH octamers in the fully compressed, inhibited conformation. But we observed two key structural differences that correlated with assembly state: the conformation of the N-terminus, and the relative orientation of protomers within each tetramer. The filament interface of the inhibited segments is unchanged from the uninhibited filaments (Figure 4F). However, at the free filament ends the N-terminus is only partially unresolved (a.a. 1–11), with the resolvable portion rotated ~30° degrees compared to the bound interface, such that Val13 inserts into the shallow hydrophobic pocket formed by A307, A308, and Lys311 of the neighboring protomer, very similar to its position in the GTP-bound free octamer crystal structure (PDB 6i0o) (Figure 4G) (Fernández-Justel et al., 2019). The resolution of the free octamer structure precludes side chain placement, but the backbone density of the unbound N-terminus is in the same position, indicating that free filament ends and free octamers are in the same conformation (Figure 4H).

We noticed a striking difference between the conformation of the catalytic core assembly in free versus assembled states. Each protomer of the free octamer and at filament ends is tilted ~5° relative to the four-fold symmetry axis, such that the tetramer becomes more ‘bowed’ than the ‘flat’ tetramers at filament assembly interfaces (Figure 4I, Video 1). The tetramers found in the GTP-bound free octamer crystal structure (PDB 6i0o) are in a similar bowed conformation (Fernández-Justel et al., 2019). The bowed tetramer conformation and the filament assembly interface are mutually exclusive. When the N-terminus conformation from the free octamer is modeled into the flat tetramer of the filament assembly interface, V13 is not correctly positioned to bind the A307/A308/K311 of the neighboring protomer, and instead clashes with K311 and A307 (Figure 4—figure supplement 6A). Thus the flat conformation promotes release of the N-termini from their binding sites on neighboring protomers, freeing them to rotate into the conformation seen in the assembled filament interface. The reciprocal operation is also not possible; for a bowed tetramer modeled with the N-terminus conformation from the filaments, only one protomer at a time can form the assembled interface (Figure 4—figure supplement 6B). The other N-termini are out of position, with significant separation of the critical residues Y12 and R356 as well as multiple steric clashes. Thus bowed tetramers cannot form the IMPDH2 filament assembly interface, and tetramer flattening is a necessary precondition for IMPDH2 filament assembly (Figure 4J).

These three structures of inhibited IMPDH2 conformations provide a model by which filament assembly is influenced by IMP and GTP through tetramer bowing and flattening (Figure 4K). In the absence of substrate, GTP induces both compression of filament segments and tetramer bowing, with the latter resulting in filament disassembly. But when IMP is bound, the disordered active site loops become ordered and rigid, buttressing intra-tetramer contacts as well as forming a pseudo-beta-barrel between opposing tetramers in the GTP-bound state. These increased contacts work to resist tetramer bowing and more readily sample the flat tetramer conformation, which promotes, and is stabilized by, the filament assembly interface. When IMP levels are low, GTP promotes filament disassembly, but high IMP levels shift the equilibrium towards filament assembly.

Based on this model, during quiescence, when the salvage pathways supply ample GTP and IMP production is downregulated, intracellular IMPDH2 will be in the fully compressed, fully inhibited, ATP/GTP-bound free octamer state. Without increased IMP production, guanine depletion will result in transient octamer extension and filament assembly, both of which will reverse as the resulting increase in IMPDH flux restores guanine levels by diverting substrate from the parallel de novo adenine pathway, mirroring the extension/compression behavior of non-filament assembling homologues (Buey et al., 2017; Fernández-Justel et al., 2019). Recently, it was shown that in vivo IMPDH assembly is promoted by increases in intracellular IMP (Keppeke et al., 2018). We therefore reasoned that the primary regulatory function of filament assembly may apply only in the proliferative state, when intracellular IMP levels are upregulated, preventing filament disassembly by GTP.

To probe whether filament assembly influences the regulatory effects of GTP on IMPDH2 activity we compared enzyme kinetics of the wildtype enzyme with the non-assembling mutant Y12A. We measured the GTP dose-response of IMPDH2 pre-incubated with varying levels of ATP, and found that wild-type enzyme is less sensitive to GTP inhibition, compared with Y12A (Figure 5—figure supplement 1A–C). Depending on ATP concentration, the apparent GTP IC50 of the wild-type was roughly two-fold lower than for Y12A. At higher ATP levels, the apparent GTP IC50 of WT and the non-assembly mutant both increase. We attribute this to competition between ATP and GTP at the first and second Bateman sites, suggesting that independent of filament assembly, GTP inhibition of IMPDH2 is affected by ATP. Notably, the range of GTP in which the substrate-saturated filaments resist GTP inhibition is within the upper range of in vivo concentrations (Traut, 1994). One function, then, of IMPDH2 filaments is to resist inhibition by GTP.

To correlate structural differences in IMPDH2 filaments as a function of GTP concentration, we collected negative stain EM data directly from three reaction volumes used for the GTP inhibition experiments, corresponding to uninhibited filaments (no GTP),~10% inhibited filaments (2.5 mM GTP), and fully inhibited filaments (20 mM GTP) (Figure 5A). As in the cryo-EM datasets, we found that uninhibited filaments were often extended, and fully inhibited filaments were universally compressed. However, we noted that the partially inhibited filaments contained a more heterogeneous mix of extended, bent, and compressed segments.

Next, we compared NAD⁺ kinetics of uninhibited (0 mM GTP) and partially inhibited (2 mM GTP) filaments with saturating IMP concentrations (Figure 5B–C). As expected, in the absence of GTP WT and the non-assembling mutant have similar apparent Michaelis-Menten kinetics (Anthony et al., 2017). Inhibition of WT by this concentration of GTP is partial and non-competitive. In contrast, the non-assembling mutant is strongly inhibited. In the absence of saturating IMP, these same GTP levels result in complete octamer compression and filament disassembly, providing a possible mechanism by which filament assembly alters GTP inhibition: by resisting the fully compressed state.

To better understand this partially inhibited state, we collected cryo-EM data of IMPDH2 with saturating substrates, 0.5 mM ATP, and 2 mM GTP (Figure 5—figure supplement 1D). Under these conditions, we observed the non-assembly mutant Y12A was 84% inhibited but WT enzyme was only 16% inhibited. We resolved structures of not only the filament assembly interface and several distinct filament segments, but surprisingly, two different free octamers; a canonical free octamer (a dimer of tetramers bound by Bateman domains), and also a small class of free interfacial octamers (a dimer of tetramers bound by the filament assembly interface) (Figure 5—figure supplement 1E, Figure 5—figure supplement 2A, Figure 5—source data 1).

The 3.1 Å resolution consensus assembly interface map is identical to the uninhibited and fully inhibited consensus interfaces, including a well-resolved active site with strong density for both substrates; as with the ATP/IMP/NAD+ structures these filaments are actively turning over and we have likely captured many states, which we have modeled simply as IMP/NAD⁺ (Figure 5—figure supplement 2B–C, Figure 5—figure supplement 5A, Figure 5—figure supplement 7A). As expected, the filament segments exhibited a range of conformations (Figure 5—figure supplement 2D). The best resolved of these were a fully compressed filament segment at 3.4 Å resolution, and two bent conformations at 4.2 and 3.7 Å resolution (Figure 5—figure supplement 2E–M, Figure 5—figure supplement 5B–D). The fully compressed filament segment is identical to the corresponding structure from the inhibited (20 mM GTP) filament dataset, with all Bateman binding sites occupied (Figure 5—figure supplement 7B). As with the bent conformation from the ATP/IMP/NAD⁺ dataset, the two bent filament segments are each two-fold symmetric octamers of asymmetric tetramers containing different proportions of extended or compressed protomers (Figure 5—figure supplement 7C–D). The asymmetric unit from one of these bent segments is a tetramer with two compressed and two extended protomers, and the other has three compressed and one extended. For each of the compressed protomers, there is clear ligand density bound at Bateman Site 3; for the extended protomers this site is unformed and empty.

Unlike the mixture of extended and compressed protomers we observed in the filaments, the free canonical octamers were universally compressed (Figure 5—figure supplement 3A–C). As with the ATP/GTP free octamer, the best resolved canonical free octamer class (3.8 Å) is fully compressed and bowed, despite having both substrate sites occupied (Figure 5—figure supplement 3D–F, Figure 5—figure supplement 6A, Figure 5—figure supplement 7E). Thus, even in the IMP-bound state, the compressed/flat conformation is less preferred, unless stabilized by the filament assembly interface. The small class of free interfacial octamers (3.8 Å) was no different from the filament interface maps, except that the Bateman domains are completely unresolved, indicating that without the stabilization provided by Bateman domain dimerization, these regions are highly flexible (Figure 5—figure supplement 4A–F, Figure 5—figure supplement 6B, Figure 5—figure supplement 7F). The observation of dramatically different structural ensembles in filament-bound and free IMPDH2, in the presence of identical ligand concentrations, explains the role of filament assembly in resisting compression and GTP inhibition.

From our different cryo-EM datasets combined, we have now determined structures of canonical IMPDH2 octamers bound to allosteric effectors and both substrates, in six distinct conformations (Figure 5D). From the uninhibited ATP/IMP/NAD⁺ dataset we resolved a fully extended filament segment, and a bent segment in which for each tetramer, 3 protomers were extended and 1 was compressed. From the partially inhibited GTP/ATP/IMP/NAD⁺ dataset, we resolved 2:2, 3:1, and fully compressed filament segments, and a fully compressed free octamer. These five structures provide a mechanism by which the Bateman domain extension of discrete protomers promoted by filament assembly resists GTP inhibition (Figure 5E). Depending on the degree of extension/bending/compression, filament segments exhibit a range of increasingly stronger interactions between opposing catalytic domains. For the fully extended segments, there are no interactions, and the flexible active loops are able to perform the complex conformational changes necessary for catalysis (Buey et al., 2017; Hedstrom, 2009). Going through the progressively more compressed bent states, there is progressively greater surface area buried by a series of distinct contacts between the opposing active site loops, until the fully compressed filament segment. For the active sites that make these contacts, activity is likely impaired because the active site loops are constrained. However, the presence of some unconstrained active sites in the bent filament segments means that these protomers are catalytically active. This effect is not cooperative within the octamers, and even a single extended protomer is sufficient to reduce inhibitory active site contacts.

The Bateman domain ligand occupancy of these five filament segment conformations varies significantly. Given the resolution range of these structures (3–4 Å) it is not possible to unequivocally distinguish ATP from GTP, and we have assigned ligand identity according to previous structures of a bacterial IMPDH bound to ATP (PDB 4dqw) (Labesse et al., 2013) and a fungal IMPDH bound to ATP and GDP (PDB 5tc3) (Buey et al., 2017). These were chosen due to conformational similarity of the Bateman domains of these structures to our ATP/IMP/NAD⁺ fully extended and ATP/IMP/GTP fully compressed human filament cryo-EM structures; backbone RMSDs were 1.163 and 0.856 Å, respectively (for residues corresponding to human IMPDH2 residues 110–244). As previously described, in both the fully extended and 1:3 compressed:extended segments there is clear ligand density at Bateman sites 1 and 2 (Figure 5F). For the single compressed protomer in the latter, site 3 is formed but due to the absence of GTP in that dataset it is unoccupied. In the partially inhibited filament dataset, for which the buffer contained both ATP and GTP, we see a greater variation in Bateman ligand occupancy. For the 2:2 compressed:extended segment, there is full occupancy at sites 1, 2, and 3 in the compressed protomers, however for the two extended protomers there is strong density at site 1 and partial density at site 2. The 3:1 compressed:extended segment is qualitatively similar; the three compressed protomers possess full ligand occupancy but for the one extended protomer site 2 has only partial density. For both the fully compressed filament segment and the free octamer there is full ligand density. Thus filament assembly promotes the extended state of individual protomers, which reduces overall GTP affinity due to disruption of site 3.

To probe whether the conformations of filament segments co-vary along the length of the filaments, we subjected the five filament-interface cryo-EM datasets presented here to classification against a series of synthetic templates with different conformations, and compared the conformations of interfaced octamers through calculations of odds ratios (Figure 5—figure supplements 8–13). The results of the classification against synthetic templates are in line with the experimentally derived classifications for each dataset, but we did not find a strong correlation in segment conformation along filaments, suggesting that conformational changes are not cooperative across the filament interface.

Purified hIMPDH2 was prepared as described previously (Anthony et al., 2017). BL21 (DE3) E. coli transformed with a pSMT3-Kan vector expressing N-terminal SMT3/SUMO-tagged hIMPDH2 were cultured in Luria broth at 37°C until reaching an OD₆₀₀ of 0.8 and then induced with 1 mM IPTG for 4 hr at 30°C and pelleted. The remainder of purification was performed at 4°C. Pellets were resuspended in lysis buffer (50 mM KPO4, 300 mM KCl, 20 mM imidazole, 800 mM urea, pH 8) and lysed with an Emulsiflex-05 homogenizer. Lysate was cleared by centrifugation and SUMO-tagged hIMPDH2 chromatagraphically purified with HisTrap FF columns (GE Healthcare Life Sciences) and an Äkta Start chromatography system. After in-column washing with lysis buffer and elution (with 50 mM KPO4, 300 mM KCl, 500 mM imidazole, pH 8), peak fractions were treated with 1 mg ULP 1 protease (Mossessova and Lima, 2000) per 100 mg hIMPDH2 for 1 hr, followed by the addition of 1 mM dithiothreitol (DTT) and 800 mM urea. Protein was then concentrated using a 30,000 MWCO Amicon filter and subjected to size-exclusion chromatography using Äkta Pure system and a Superose 6 column pre-equilibrated in filtration buffer (20 mM HEPES, 100 mM KCl, 800 mM urea, 1 mM DTT, pH 8). Peak fractions were flash-frozen in liquid nitrogen and stored at −80°C.

Filaments (or, depending on ligand state, free octamers) were prepared by diluting aliquots of purified hIMPDH2 in activity buffer (20 mM HEPES, 100 mM KCl, 1 mM DTT, pH 7.5) to 2 μM in the presence of varying concentrations of ATP, GTP, IMP, and/or NAD⁺ and incubating for 30 min at 20°C. Nucleotide stocks were prepared using ATP disodium salt hydrate (Sigma A2383), GTP sodium salt hydrate (Sigma G8877), IMP disodium salt hydrate (Acros AC226260050), and β-Nicotinamide adenine dinucleotide hydrate (Sigma N6522), For all buffers, protein preparations, and nucleotide stocks, final pH was checked and titrated if necessary to 7.5 with KOH or HCl.

Protein aliquots were diluted in activity buffer and pre-treated with varying concentrations of ATP, GTP, and IMP for 30 min at 20°C, in 96 well UV transparent plates (Corning model 3635). Reactions (100 uL total) were initiated by addition of varying concentrations of NAD⁺. NADH production was measured by optical absorbance (340 nm) in real-time using a Varioskan Lux microplate reader (Thermo Scientific) at 20°C, 1 measurement/min, for 20 min; absorbance was correlated with NADH concentration using a standard curve. Specific activity was calculated by linear interpretation of the reaction slope for a 5 min window beginning 3 min after reaction initiation.

Protein preparations were applied to glow-discharged continuous carbon EM grids and negatively stained with 2% uranyl formate. Grids were imaged by transmission electron microscopy using an FEI Tecnai G2 Spirit at 120kV acceleration voltage and a Gatan Ultrascan 4000 CCD using the Leginon software package (Suloway et al., 2009). Micrographs were collected at a nominal 67,000x magnification (pixel size 1.6 Å). GCTF was used for contrast transfer function (CTF) estimation, and Relion for particle picking and 2D classification (Scheres, 2012; Zhang, 2016; Zivanov et al., 2018).

Protein preparations were applied to glow-discharged C-flat holey carbon EM grids (Protochips), blotted, and plunge-frozen in liquid ethane using an Vitrobot plunging apparatus (FEI) at 20°C, 100% relative humidity. High-throughput data collection was performed using an FEI Titan Krios transmission electron microscope operating at 300 kV and equipped with a Gatan image filter (GIF) and post-GIF Gatan K2 Summit direct electron detector using the Leginon software package (Suloway et al., 2009). For the two datasets with non-filament octamers of IMPDH, which exhibit a preferred orientation, it was necessary to collect images with the stage tilted in order to capture a sufficient range of views for 3D reconstruction (Figure 4—figure supplement 3A, Figure 5—figure supplement 2A).

Movies were collected in super-resolution mode, then aligned and corrected for beam-induced motion using Motioncor2, with 2X Fourier binning and dose compensation applied during motion correction (Suloway et al., 2009; Zheng et al., 2017). CTF was estimated using GCTF (Zhang, 2016). Relion 3.0 was used for all subsequent image processing (Zhang, 2016; Zivanov et al., 2018). Although multiple datasets of hIMPDH2 under the different ligand states were collected, each dataset was individually processed using approximately the same overall pipeline (Figure 2—figure supplement 1E–G), with some variations from dataset to dataset (Figure 3—figure supplement 1, Figure 4—figure supplement 2, Figure 4—figure supplement 3, Figure 5—figure supplement 2, Figure 5—figure supplement 3, Figure 5—figure supplement 4). First, for each dataset, autopicking templates and initial 3D references maps were prepared by manually picking and extracting boxed particles from a small subset of micrographs, and classifying/refining in 2D and 3D. For these initial 3D refinements, a featureless, soft-edged cylinder was used as a refinement template of filaments, and a previously published cryo-em map (EMDB-8692) was used as template for non-filament octamers (Anthony et al., 2017). Because IMPDH filament segments possess D4 point-group symmetry, two different locations along filaments may be used as symmetry origins: the centers of canonical octamer segments, or the centers of the assembly interface between segments. For the filament datasets, we prepared and used auto-picking templates centered on the filament assembly interface. For the datasets containing non-filament octamers of hIMPDH2, auto-picking templates centered on these non-filament particles were also included. Due to the flexibility of hIMPDH2 filaments, helical segments were processed as single particles, and at no point was helical symmetry applied during image processing. After template-based autopicking of each complete dataset, picked particles were boxed and extracted from micrographs, and subjected to hierarchical 2D classification to select the best-resolved classes. These selected particles were then auto-refined in 3D as a single class with symmetry applied (D4 for filament segments and free octamers, C4 for filament ends). Exploratory image processing of the assembly interface-centered filament reconstructions made it apparent that the eight catalytic domains surrounding this interface appeared conformationally homogenous, while the Bateman domains and neighboring octamers appeared conformationally varied. Additionally, due to the flexibility of the filaments, and the tendency of filament ends to adhere to the air-water interface, many filaments were tilted out of plane, with neighboring segments overlapping in projection.

To improve resolution, partial signal subtraction was performed at this stage, using a mask that left only the central eight catalytic domains of the filament assembly interface, subtracting the poorly resolved Bateman domains and neighboring segments, which served to improve resolution after subsequent auto-refinement. Per-particle defocus and per-micrograph astigmatism were then optimized using CTF refinement, which improved resolution further. The resulting consensus refinements of the filament assembly interfaces were well-resolved, however data on Bateman domain conformation was missing, with these regions very poorly resolved when subtracted regions were restored to the reconstructions by reversion to original non-subtracted particles (data not shown). To resolve the different Bateman domain conformations, we applied particle symmetry expansion (D4 to C1) and classified particles without additional alignment. Because at this stage the reconstructions were centered on the filament assembly interface, each boxed ‘particle’ contained elements of two different neighboring octamers. The potential conformational space was reduced by applying a mask enclosing only one of these two octamers. By hierarchical focused classification of the off-origin octamers we were able to classify multiple conformations of the octameric filament segments, as well as incomplete segments and filament ends. Symmetry expansion was also applied to the non-filament octamer datasets, with a mask including the entire particle, which allowed classification of the most symmetric and well-resolved classes. To further improve resolution of the varying symmetry-expanded segment classes, the reconstruction symmetry origins were moved from the filament assembly interface to the canonical octamers by re-extraction with re-centering. For each class, symmetry was then collapsed by removing redundant overlapping particles, Euler angles reset to zero. After auto-refining once again, we observed that the most well-resolved octameric segments from the asymmetric symmetry-expanded classifications exhibited some apparent symmetry, with fully extended or fully compressed octamers appearing D4 symmetric, and some bent classes apparently D1 symmetric. We therefore applied these symmetries during subsequent refinement and classification of these new octamer-centered classes. As before, signal subtraction of neighboring filament segments improved resolution considerably. Additional rounds of CTF refinement and 3D classification identified the best-resolved particles from each of these conformational classes. Final overall resolution (according to the FSC = 0.143 criterion), as well as local resolution, was assessed using Relion postprocessing.

As initial templates for model building, two hybrid models (representing hIMPDH2 in either an extended or compressed state) were prepared by combining elements from existing crystal structures. For both templates, the catalytic domain and substrate poses (residues 18–107, and 245–514) were taken from a crystal structure of an inhibitor-bound hIMPDH2 (PDB 1nf7), and the Bateman domains and ligand poses (residues 108–244) were based on fungal (A. gossypii) IMPDH crystallized in either the extended or compressed states (PDB 5mcp and 5tc3, respectively), and SWISS-MODEL homology modeling (Buey et al., 2017; Sintchak et al., 1996; Waterhouse et al., 2018). The N-terminus (residues 1–17) were modeled by hand. In all maps, a single active site loop (residues 421 to 436) was unresolved, and these residues were not modeled. After rigid-body fitting of templates into the cryo-EM densities using UCSF Chimera, repeated cycles of manual fitting with Coot, automated fitting with phenix.real_space_refine (employing rigid-body refinement, NCS constraints, gradient-driven minimization and simulated annealing) and local B-factor sharpening of cryo-EM data via LocScale were used for final atomic model refinement and local sharpening of cryo-EM maps (Adams et al., 2012; Emsley et al., 2010; Jakobi et al., 2017; Pettersen et al., 2004). Final models were evaluated with MOLPROBITY and EMRinger (Barad et al., 2015; Chen et al., 2010). Data collection parameters and refinement statistics are summarized in Figure 3—source data 1, Figure 3—source data 2, Figure 3—source data 3, Figure 4—source data 1, Figure 4—source data 2, Figure 5—source data 1. Figures were prepared with UCSF Chimera (Pettersen et al., 2004). Fourier shell correlations between the final models and maps were calculated with phenix.mtriage, and to probe for overfitting, a model refined against one half-map was likewise used to calculate FSCs between both that half-map (FSC-work) and the other half-map (FSC-test) (Afonine et al., 2018) (Figure 3—figure supplement 2, Figure 4—figure supplement 3, Figure 4—figure supplement 5, Figure 5—figure supplement 5, Figure 5—figure supplement 6).

The five interface-centered filament particle stacks were re-extracted to a larger box size and a common pixel size (2 Å/px) and a larger box size (576 Å) that completely included both interfaced octamers (Figure 5—figure supplement 8A). Synthetic templates representing different octamer conformations and orientations for either of these interfaced octamers were prepared using a subset of the final experimental models, and simulating EM density using the Eman2 tool pdb2mrc (Tang et al., 2007) (Figure 5—figure supplement 8B). The paired octamers from each interface were then classified in Relion for a single iteration, without additional alignment, using masks to focus on one octamer or the other. Odds ratios (and associated 95% confidence intervals) were then calculated for all possible pairs of classified conformations between the sets of interfaced octamers ((Figure 5—figure supplement 8C, (Figure 5—figure supplements 9–13).