Structural insights into the functional cycle of the ATPase module of the 26S proteasome

To obtain a high-resolution 3D map, images of vitrified yeast 26S proteasomes in the presence of 4 mM ATP were recorded (Fig. S1A and Table S1). After elimination of low-quality particles (Fig. S1 B and C), the proteasome particles were classified according to the RP states (22, 26). Ultimately, we obtained maps of the s1 and s2 states to global resolutions of 4.1 Å and 4.5 Å, respectively (Fig. 1 A and B and Fig. S1F). Similar to the human 26S proteasome (22), the CP exhibited local resolution values beyond 1/4 Å⁻¹, whereas the RP, especially the peripheral regions, was resolved at lower than 1/8 Å⁻¹ (Fig. S1 G and H).

We first built the model of the s1 state from a comparative model based on the human 26S proteasome structure (22). The s2 model was obtained by molecular dynamics flexible fitting (MDFF) (27) of the s1 model into the s2 map. The s1 and s2 structures are overall similar to the previous models based on low-resolution densities (21), but they exhibit subtle structural differences, such as the register shifts as described by Schweitzer et al. (22). These differences include a kink in the 26S proteasome ATPase regulatory subunit 1/2 (Rpt1/2) N-terminal coiled coil (Rpt1 Gly68 to Pro78, Rpt2 Ile81 to Pro84), which was not found in the other two coiled coils (Rpt3/6 and Rpt4/5).

Despite the overall lower resolution, Rpn13 is better resolved in the s2 state than in the s1 state. The structure and motion of this subunit were analyzed further by focused classification. An exhaustive 6D correlation search in the best-resolved class (Fig. S2A) resulted in an unambiguous positioning of Rpn13. Cross-linking mass spectrometry (MS) data validated this positioning (Table S2). In comparison to the previous model (21), Rpn13 is rotated by ∼105°, so that the Rpn2-binding region of Rpn13 points toward Rpn2 (28) and fulfills the cross-linking restraints.

To detect additional conformations of the RP, we analyzed structures of the 26S proteasome in the presence of different nucleotides and nucleotide analogs [adenylyl-imidodiphosphate (AMP-PNP), ATP/ADP + beryllium fluoride (BeF_x), and ADP] (Table S1). AMP-PNP and ADP-BeF_x are known as “nonhydrolyzable” nucleotide analogs that mimic an ATP ground state with tetrahedral geometry similar to the geometry of the “slowly hydrolyzable” nucleotide analog adenosine 5′-[γ-thio]triphosphate (ATP-γ-S) (29). Upon 3D classification, each dataset showed different abundances of the conformational states (Fig. S3B). Although the inhibition of the ATPase activity by AMP-PNP was weaker than for other analogs (Fig. S3A), the class averages of the AMP-PNP dataset showed exclusively the s3 conformation. In contrast, the presence of ADP alone led to highly heterogeneous conformations, which could not be classified into distinct states (Table S1). In the BeF_x dataset, particles were distributed into four different classes (Fig. S3B). One of these classes showed a previously unobserved conformation of the 26S proteasome (Fig. 2B and Fig. S4), which we termed the s4 state. In the ADP-BeF_x and ATP/BeF_x samples, the occupancy of the s4 class was, respectively, 2% and 9%. After angular refinement, the s3 and s4 maps yielded resolutions of 7.8 Å and 7.7 Å, respectively (Fig. 2 A and B and Fig. S1F).

We modeled the s3 structure using MDFF to fit the s2 model into the s3 density, and the s4 structure based on the resulting s3 model (Fig. S3C). Secondary structure elements like α-helices were detected and positioned for all subunits. However, the low-resolution EM densities of the s3 and s4 states did not allow for the unambiguous identification of side chains.

Although, overall, the s4 structure seems similar to the s3 structure, some subunits, such as Rpn1, Rpn6, Rpn11, and the AAA⁺ ATPase, adopt different conformations (Fig. 2C). For example, compared with the s3 conformer, the N-terminal α-solenoid of Rpn6 moves by ∼12 Å toward Rpn7, independent of the overall rotation of the lid complex, whereas Rpn11 shifts by ∼11 Å together with the entire oligosaccharide-binding (OB) ring. The rigid-body rotation of Rpn1 from s3 to s4 positions the Rpn1 N terminus ∼24 Å closer to the AAA⁺ ring (Fig. S3 E and F). Compared with the s1 state, Rpn1 shifts and rotates in the s4 state, leading to an interaction of its N terminus with Rpt2 and Rpt6. This conformational change increases the distance between the central leucine-rich repeats domain of Rpn1 and the coiled coil of Rpt4/5, which might facilitate Ubp6 binding (Fig. S3F). Indeed, we observed an additional density between Rpn1 and the OB ring in the s4 state (Fig. 2B, box), which coincides with the position of Ubp6 (30). Further focused classification of the s4 dataset revealed that ∼50% of the particles do not possess Ubp6 (Fig. S3D), roughly consistent with the amount of Ubp6 in the sample determined by MS.

The diameter of the CP gate varies in the four observed conformations (Fig. S5). Consistent with previous studies (17, 26), the EM map of the CP in the s1 state shows high densities corresponding to the N termini of the α-subunits at the CP gate (Fig. S5 A and B). Interestingly, we observed no clear density within the central pore in the EM map of the s4 state (Fig. 2B, Inset, and Fig. S5 A and C), suggesting the opening of the CP gate. To quantify gate opening, we calculated the radial average of the EM density perpendicular to the CP axis (Fig. S5D). The normalized intensity of the s1 state is similar to the normalized intensity of the crystal structure of the CP with a closed gate [Protein Data Bank (PDB) ID code 5cz4 (31)] (Fig. S5 E and G). The s4 state shows the lowest normalized intensity at the center (Fig. S5K), in good agreement with the open-gate CP crystal structure [PDB ID code 1z7q (32)] (Fig. S5M). In the s4 model, several residues corresponding to the EM densities of the N-terminal extension are extended toward the RP (Fig. S5C), similar to the conformation observed in the CP–PA26 complex (32). The normalized intensities of s2 and s3 lie in between those extrema, indicating a partially closed gate (Fig. S5 H–J and M).

Next, we analyzed the density at the interface between the CP and RP. In contrast to the human 26S structures (22), all three HbYX motifs (Rpt2, Rpt3, and Rpt5) show clear densities inside the α-pockets in the four yeast conformational states (Fig. S6A). In the s1 structure, the conformation of the Rpt2 HbYX motif is similar to that of the HbYX motif of the proteasome-activating nucleotidase PAN in complex with the CP (11). The conserved tyrosine in the Rpt2 HbYX motif is in proximity to the arginine residue of the α4 subunit (Fig. S6B). In addition to the association of the HbYX motifs, in the s2 state, the Rpt1 C-terminal α-helix extends toward the C terminus of the α6 subunit (Fig. S6C). In the s4 state, in addition to the three HbYX motifs, the C-terminal tail of Rpt6 was detected at the interface between the α3 and α4 subunits (Fig. S6A).

In the s1 AAA⁺ ATPase, the OB and AAA⁺ rings are stacked on top of each other, but only the AAA⁺ ring forms a “lockwasher” conformation, in which the helical-shaped ring is split at the interface between Rpt6 and Rpt3 (17, 29). The large domains of the Rpt subunits are positioned at different heights. Rpt2 positions at the bottom and Rpt3 at the top (Fig. S7). The AAA⁺ ring is further right-handed twisted, arranging the α8 helices, which are C-terminal of the pore-1 loop in different orientations: Rpt2 and Rpt6 tilt inward up to 48°, and Rpt3 is almost planar with respect to the CP plane (Fig. 3). In this arrangement, the Rpt2 pore-1 loop is located closest to the CP and the Rpt3 pore-1 loop is proximal to the OB ring (Fig. 3 and Fig. S7).

Compared with the s1, in the s2 state the AAA⁺ ATPase shifts toward higher coaxial alignment with the CP, without significant local conformational changes (Fig. 3 and Figs. S4 and S7–S9), indicating that the coaxial alignment is not an ATP-driven conformational change. The AAA⁺ ring is rotated by ∼8° and moved by ∼17 Å for further coaxial alignment in the s3 state (Fig. S4). The large movement of the Rpt subunits relocates the opening of the lockwasher conformation to a site opposite of Rpt3/6, the interface of Rpt5/1. In this conformation, the α8 helices of Rpt1 and Rpt2 are tilted outward and Rpt4 and Rpt5 inward, resulting in a rearrangement of the pore-1 loop with Rpt5 at the bottom and Rpt1, Rpt2, and Rpt6 at the top (Fig. 3). Finally, the s4 AAA⁺ ring shows only a slight shift from the s3 conformation (Fig. S4). However, the lockwasher conformation, which splits at the interfaces between Rpt3/4 and Rpt4/5, is still preserved. In this conformation, the N termini of the α8 helices locate such that Rpt5 is at the top position and Rpt3 is at the bottom position. In all four conformational states, the AAA⁺ ATPase adopts different lockwasher conformations, which change the height of each pore-1 loop by switching the opening of the AAA⁺ ring.

Like other AAA⁺ ATPase family members, the nucleotide-binding pocket is formed by five conserved motifs: Walker A, Walker B, sensor I, sensor II, and Arg-finger. The nucleotide is located at the “interface module,” defined by a small domain of one Rpt subunit and the large domain of its clockwise adjacent Rpt subunit (20, 33). In the s1 state, the interface module units are closely connected to each other, except at the split site between Rpt6 and Rpt3 (Fig. 4D). Similar to previous studies (22, 24, 25), the s1 map shows density for nucleotides in all six nucleotide-binding pockets (Fig. 4 A, F, and G). However, only five of the six Arg-fingers are engaged in nucleotide binding (“engaged pocket”), in which the phosphates (Pi) of the nucleotide interact with the side chains of the two Arg residues of the Arg-finger (Fig. 4A). In the engaged pockets, the distance between the N-terminal tip of the α6 helix at the Walker A motif and the center of α10 helix located after the walker B motif (“pocket distance”) is ∼4–5 Å, showing tight ATP binding (Fig. 4 C and D). In contrast, the pocket distance of the interface module Rpt6/3 is ∼30 Å (Fig. 4D). Interestingly, the Rpt3 Arg-fingers are projected away from the nucleotide-binding pocket and form hydrogen bonds with the carboxy groups of Rpt6, Glu140, and Val141 (“open pocket”) (Fig. 4 A and C). Despite nonengagement of the Rpt3 Arg-finger, there is still nucleotide density in the Rpt6/3 pocket. Indeed, the density volume of the Rpt6/3 pocket is smaller than the nucleotide densities of the rest of the subunits, implying ADP binding (22). The conformations of the binding pockets in the s2 state are identical to the conformations of the binding pockets in the s1 state (Fig. 4 B and G), supporting the notion that there is no ATP hydrolysis involved in the conformational change from s1 to s2 (Figs. S8 and S9). Interestingly, we identified nucleotide densities in all Rpt subunits in both the s3 and s4 states (Fig. 4G). Similar to the s1/s2 states, the interface modules of the s3 and s4 states are in close contact with each other, except for the split sites, probably for tight ATP binding (Fig. 4 A–D). In the s3 state, the pocket distance of Rpt5/1, which is localized at the split site, is longer than the others (∼28 Å), indicating a different coordination of the nucleotide from the coordination in the engaged pockets. In the s4 state, the pockets of two subunits, Rpt3/4 and Rpt4/5, are expanded (∼21 Å) (Fig. 4D). Considering that the pocket distance probably corresponds to different nucleotide binding, the conformational changes between s2, s3, and s4 may reflect different nucleotide binding by the AAA⁺ ATPase.

The AAA⁺ ATPase family belongs to an additional strand catalytic glutamate (ASCE) superfamily, which includes RecA-like ATPases (33). Despite the high degree of structural conservation among the members of the ASCE superfamily (33), there are major differences in their hexameric assemblies. ClpX assembles into a closed planar ring but breaks symmetry by rotating the hinge region between the small and large domains of a subunit to form an “unloadable” conformation (34). The RecA ATPase DNA helicases, Rho and E1, also have a planar conformation, whereas DnaB₆ is assembled into a lockwasher conformation (35). In the yeast 26S AAA⁺ ATPase, all four structures adopt a lockwasher conformation with different split sites. So far, the only planar conformation of the 26S proteasome was found in the presence of a model ubiquitylated substrate, in which the pore-1 loops were suggested to be arranged in a plane (36). In the substrate-engaged structure, the position of the AAA⁺ ring with respect to the CP is similar to the positions of the s3 and s4 conformers. However, the conformation of each Rpt subunit, including the arrangement of the pore-1 loop, is rather close to the s3 state. In the lockwasher conformation, the elongated pitch of the pore-1 loop may allow the ATPase to translocate polypeptides with a larger distance than the planar conformation. The distance between the top and bottom positions of the pore-1 loops of the 26S proteasome is ∼2 nm, almost double the ClpX substrate translocation step size (37). This finding indicates that the proteasome pore-1 loops in the staircase arrangement can likely translocate a polypeptide with a similar or even larger step size than ClpX.

In all reported conformations of the 26S proteasome, including the “open-gate” conformation (S_D conformation) (22–25), the AAA⁺ ATPase adopts a lockwasher conformation. However, we note that the lockwasher conformation of the AAA⁺ ATPase and the staircase arrangement of the pore-1 loops in the s4 state are not similar to the S_D conformation or any other reported conformation.

A previous structural study showed that insertion of the C-terminal HbYX motifs causes a conformational change for the CP gate opening concomitant with a rotation of the α-subunits (11). Although our s4 map shows no clear density in the pore region, probably representing an open-gate state, we did not observe any significant rotation of the α-subunits: Only the N-terminal extensions turn up toward the RP, as seen in the PA26–CP complex (38). In contrast, a strong density within the gate in the s1 state (Fig. S5A) suggests a closed gate. In the s2 and s3 states, the normalized intensity within the gate is between the normalized intensities of s1 and s4, indicating incomplete gate opening. In the crystal structure of the CP with the proteasome activator Blm10, the N-terminal extensions were not in the closed conformation but did not show a clear density, suggesting a “partial closed-gate” (10). The gate conformation of the s2 and s3 states may be similar to such a partial closed-gate state. Considering the elasticity of an unfolded polypeptide (39), the partial closed-gate may hinder the diffusion of the translocated polypeptide by gripping it with the N-terminal extensions. In addition, the fact that all four states exhibit densities of three HbYX motifs in the α-pockets but only the s4 conformation adopts an open-gate conformation, suggests that HbYX engagement is not sufficient to open the gate for the yeast 26S proteasome.

The reported S_D conformation of the human 26S proteasome showed a reduced density within the gate. However, analysis of the normalized intensity of the S_D conformation (Fig. S5L) shows that the gate area is rather similar to the s3 state (Fig. S5O), implying that the S_D state represents a conformation different from the s4 state. In the S_D conformation, two additional C-terminal tails (Rpt1 and Rpt6), together with insertion of the three HbYX motifs into the α-pockets, were suggested to trigger CP gate opening (25). However, only the C-terminal tail of Rpt6 was detected in the map of the s4 state (Fig. S6A). Due to the relatively low resolution of the s4 and S_D states, it is unclear whether this difference is due to a different gate-opening mechanism in yeast and humans or to different interpretation of the EM maps.

Mechanochemical studies of ClpX have revealed that a conformational change of the pore-1 loop is coupled with Pi release to power translocation (39). In addition, among the six protomers, a subunit whose pore-1 loop is in direct contact with a translocated polypeptide may have a higher probability of hydrolyzing ATP first (40), indicating that an Rpt subunit at the top of the staircase may hydrolyze first. Interestingly, the pore-1 loop of the Rpt subunit at the top of the staircase is adjacent to the Rpt subunit with an open nucleotide-binding pocket, in which the pore-1 loop is located at the bottom of the staircase.

Based on our structures, we propose the following functional model of the 26S proteasome (Fig. S10). In the absence of substrate, the proteasome is present in the ground state (s1), which represents the lowest energy conformation of the four states. When the proteasome is activated, most likely by substrate binding, it undergoes a conformational change from the ground state to the translocation-competent states (s2–s4). Because no structural change is observed in the AAA⁺ ring between s1 and s2, this transition is most likely not ATP-dependent. The translocation of substrates may be initiated by Rpt3, whose pore-1 loop is at the top position of the staircase in the s2 state. Polypeptide binding to the Rpt3 pore-1 loop may trigger ATP hydrolysis, probably followed by Pi release. The Pi release from the Rpt3/4 pocket brings the Rpt3 pore-1 loop to the bottom of the staircase. As a consequence, the pockets of the neighboring subunits (Rpt3/4 and Rpt4/5) undergo a conformational change, resulting in the relocation of the Rpt5 pore-1 loop at the top position (s4). A concerted conformational change occurs to open the CP gate, which allows substrate translocation into the CP antechamber. Coupling the gate opening with ATP hydrolysis prevents the polypeptide from slipping back. In a similar manner, the Rpt5 pore-1 loop undergoes a conformational change coupled with Pi release to be located at the bottom of the staircase (s3), again powering the work for the translocation. It is very likely that there exists more states than the four discussed here, in which the nucleotide pockets of Rpt1/2 and 2/6 are open. We assume that the cycle continues until the substrate translocation process is finished. Our structures favor a model in which the hydrolysis cycle occurs in a sequential order around the ring rather than in a stochastic manner.

Yeast 26S proteasomes were isolated from Saccharomyces cerevisae by affinity purification using the 3× FLAG-tagged subunit Rpn11. Different ATP analogs were introduced during or after the sucrose gradient purification step (SI Materials and Methods). Samples (∼1 mg/mL) were stored at −80 °C until further use.

All datasets were collected on a Titan Krios with a direct electron detector in movie mode. The pixel size at the specimen level was 1.35 Å for Falcon cameras and 1.38 Å for the K2 camera (SI Materials and Methods). Both single-capped 26S and double-capped particles were used in all datasets for classification and to obtain the final reconstructions (SI Materials and Methods). All image processing steps were carried out in TOM (41) and RELION (42).

Initial models were obtained by comparative and de novo modeling (SI Materials and Methods). The initial structure for the s1 state was the merged structure of the CP crystal structure [PDB ID code 5cz4 (31)] and the RP homology model based on the human structure [PDB ID code 5l4g (22)]. The subunits were positioned into the EM map and subsequently refined. Real-space refinement was first performed as described by Goh et al. (43) using MDFF (27) and then in reciprocal space. MDFF simulations were prepared using QwikMD (44), analyzed with VMD (45), and carried out with NAMD (46). The refined final structure of s1 was used to initiate an MDFF run into the density of the s2 state. The final refined structure of the s2 state was fitted through MDFF into the density of the s3 state, and the final s3 state structure was finally fitted into the s4 density.

The 26S proteasome was purified by affinity purification using the 3× FLAG-tagged subunit Rpn11. Saccharomyces cerevisiae cells (YYS40: MATa RPN11-3FLAG::HIS3) were grown for 48 h in YPD medium at 30 °C. The cells were harvested by centrifugation, and the pellet was washed with water. The yeast cell pellet was diluted with buffer A [50 mM Tris⋅HCl (pH 7.4), 100 mM NaCl, 10% (vol/vol) glycerol, 10 mM MgCl₂, 4 mM ATP, 16 mM creatine phosphate, 0.03 mg/mL creatine phosphate kinase] in a ratio of 1 g of cells to 1 mL of buffer. Cells were disrupted using Zirkonia glass beads with a glass bead mill. The supernatant was separated from the beads and cell debris by subsequent centrifugation steps. The pH of the crude extract was adjusted to pH 7.4 using Tris base. FLAG-tagged proteasomes were purified by immunoprecipitation using M2 anti-FLAG beads (Sigma). The filtered crude extract was incubated with the beads for 1.5 h at 4 °C (1 mL of resin per 40 mL of crude extract). Unbound proteins were separated by washing with buffer A. The proteasome was eluted using 200 μg/mL 3× FLAG-peptide in buffer A. The eluate was concentrated by centrifugation in Amicon spin columns (10-kDa molecular weight cut-off). To purify double-capped proteasomes further, the concentrated eluate was applied to a sucrose gradient [15–45% sucrose (wt/vol), 20 mM Hepes (pH 7.4), 40 mM NaCl, 5 mM DTT, 10 mM MgCl₂, 4 mM ATP, 16 mM creatine phosphate, 0.03 mg/mL creatine phosphate kinase] and centrifuged for 17 h at 4 °C at 208,000 × g in a Beckman SW41 rotor. To obtain proteasome samples with different ATP analogs, ATP and the ATP regenerating system of the sucrose gradient were substituted with 2 mM ADP, 2 mM adenylyl-imidodiphosphate (AMP-PNP), or 2 mM ADP/2 mM beryllium fluoride (BeF_x), respectively. BeF_x was obtained by mixing NaF with BeCl₂ [2:1 (wt/wt)]. To get proteasomes with ATP and BeF_x, the proteasome sample containing ATP was incubated with 5 mM BeF_x after sucrose gradient centrifugation. Proteasome-containing fractions were detected by the degradation of the peptide N-succinyl-Leu-Leu-Val-Tyr-7-amino-4-methylcoumarin, SDS/PAGE analysis, Bradford assay, and negative-stain EM. The fraction with the highest specific activity was analyzed by MS to estimate the subunit abundances using label-free, intensity-based absolute quantification (47). The samples were stored at −80 °C after quickly freezing with liquid nitrogen.

To estimate the inhibition of the proteasomale AAA⁺ ATPase by different ATP analogs, the hydrolysis rate of ATP was measured using the colorimetric assay described by Bartolommei et al. (48). One hundred microliters of proteasome sample (20 μM), which was purified in the presence of ADP, was incubated with 900 μL of color solution [125 mM sulfuric acid, 0.5 mM ammonium molybdate tetrahydrate, 10 mM ascorbic acid, 0.04 mM potassium antimony(III) tartrate hydrate] for 35 min at room temperature, and the OD₈₅₀ was determined. In the first step, the measurement was calibrated using different KPO₄ concentrations from 0 to 500 μM in 50 mM Hepes (pH 7.4) and 10% (vol/vol) glycerol. The proteasome ATPase activity was measured in ATP-containing reaction buffer [50 mM Hepes (pH 7.4), 10 mM MgCl₂, 4 mM ATP, 10% (vol/vol) glycerol] over 90 min at 30 °C. The phosphate concentration was determined at different time points using the same method as for the KPO₄ calibration curve. The inhibition of the AAA⁺ ATPase by BeF_x was determined by measuring the ATP hydrolysis rate after addition of BeF_x to the reaction buffer. For AMP-PNP and adenosine 5′-[γ-thio]triphosphate tetralithium salt (ATP-γ-S), the ATPase activity was measured by using the sample purified in the presence of 2 mM AMP-PNP or 2 mM ATP-γ-S.

Data acquisition was performed with an FEI Titan Krios electron microscope. Proteasome samples were plunge-frozen on Quantifoil 2/1 or Lacey carbon-coated grids using a Vitrobot MK IV instrument. Datasets were collected with a Falcon II, Falcon III, or K2 camera using either FEI EPU or the in-house developed software TOM² (49). Movies were acquired at a pixel size of 1.35 Å at the specimen level for datasets collected with a Falcon camera and 1.38 Å for datasets collected with a K2 camera. A total dose of ∼45 electrons was distributed over 7 frames for the Falcon II camera, 24 frames for the Falcon III camera, and 20 frames for the K2 camera. The nominal defocus range of the acquisition varied from 1.5 to 3.5 μm. Movies acquired with the K2 camera were corrected for anisotropic magnification distortion using the software “mag_distortion_correct” (50).

Image processing was performed similar to the procedure described by Aufderheide et al. (30). In brief, all movie frames were aligned translationally and summed with an in-house implementation of the algorithm from the study by Li et al. (51). Next, the contrast transfer function (CTF) was estimated using CTFFIND3 (52), and micrographs with a defocus outside the range of 0.8–3.5 μm and a CTF fit score below 0.05 were discarded. After micrograph inspection, 66,462 micrographs of the ATP dataset (K2), 3,304 micrographs of the AMP-PNP dataset (Falcon II), 11,121 micrographs of the ATP-BeF_x dataset (Falcon II), 6,522 micrographs of the ADP-BeF_x dataset (Falcon III), and 5,732 micrographs (Falcon II) plus 6,136 micrographs (Falcon III) of the ADP dataset remained.

Single-particle analysis was performed as described by Aufderheide et al. (30). In the first step, micrographs were subjected to automated localization of 26S proteasome particles using the TOM toolbox as described by Beck et al. (26). All further single-particle analysis steps were performed using the RELION software package (42). Proteasome particles were extracted with a box size of 384 × 384 for K2 datasets and 416 × 416 for Falcon datasets, resized to a box size of 256 × 256, and subjected to reference-free 2D classification. Only 2D classes containing high-quality particles were retained. Additionally, particles were classified into single-capped 26S (sc26S) and double-capped 26S (dc26S) proteasome particles according to their 2D class averages.

For the ATP dataset, 3D structures of 419,464 sc26S and 269,854 dc26S particles were reconstructed with the RELION software package, using a down-filtered 3D reference of the 26S proteasome, reconstructed from particles from Unverdorben et al. (21) and filtered to 60 Å. Similar to a previous proteasome study (22), the reconstruction result indicated an uneven angular distribution (Fig. S1 C and D). To reduce the size of the dataset, angular classes with an above-average occupancy were reduced to the mean occupancy by discarding those particles that score worst in terms of the _rlnMaxValueProbDistribution value in RELION, as described in Schweitzer et al. (22). The remaining 196,475 sc26S and 194,803 dc26S particles were subjected to particle polishing. Polished sc26S (C1 symmetry) and dc26S (C2 symmetry) particles were refined using RELION auto-refinement. After 3D alignment of the polished sc26S particles, these particles were classified into 10 classes using a soft-edged mask focusing on the RP and keeping the previously assigned angles constant. In line, polished double-capped particles were aligned, and after the broken C2 symmetry was addressed, the RPs of the dc26S (pseudo-sc26S) particles were classified into 10 classes using a soft-edged mask on the RP and keeping the previously assigned angles constant. The 3D class averages were compared with the previously identified proteasome states s1, s2, and s3 (21) using the UCSF Chimera (53) fit-in map. Among the sc26S particles, 83,007 were classified into classes showing the s1 state of the proteasome and 50,066 showing the s2 state. Particles from 3D class averages that did not reconstruct into a well-defined 26S proteasome were discarded (“broken” particles). Pseudo-sc26S particle classification resulted in 203,527 s1-like particles for the s1 state and 143,271 s2-like particles. The sc26S and pseudo-sc26S particles from the s1 state were combined and refined, using a soft-edged mask containing one RP and the CP, with a local angular search around the initial angles from the refinement of the polished particles. The resulting density was subjected to postprocessing in RELION for resolution determination and B-factor sharpening. Applying the mask used for refinement, an average resolution of 4.1 Å was obtained. The s2-like particles were refined and postprocessed applying the same procedure as for the s1-like particles, leading to a final reconstruction with an average resolution of 4.5 Å.

A total of 20,325 sc26S and 119,023 dc26S particles from the 2D classification of the ATP-BeF_x dataset and 75,323 sc26S and 109,021 dc26S particles from the 2D classification of the ADP-BeF_x dataset were processed according to the procedure described for the ATP dataset, including 3D reconstruction, particle polishing, alignment of sc26S and pseudo-sc26S, and classification with a mask on the RP. In contrast to the ATP dataset, the BeF_x datasets were not reduced by removing particles from angular classes with an above-average occupancy because the particle numbers were small enough to be processed by RELION. The classification into 12 classes of the ATP-BeF_x particles resulted in 52,335 s1-like and 69,397 s2-like particles. Additionally, 60,819 particles adopted a conformation that could not be assigned to the previously known states. This state is referred to as s4. In a similar manner, the particles from the ADP-BeF_x dataset were classified into 121,560 s1-like particles, 8,217 s2-like particles, 63,957 s3-like particles, and 38,661 s4-like particles. To analyze the conformational heterogeneity of the datasets further, s2-like, s3-like, and s4-like particles were combined, aligned, and further classified into 12 classes, again keeping the previously assigned angles constant. The final classification resulted in 125,903 s2-like particles, 69,723 s3-like particles, and 27,625 well-defined s4-like particles. The s4-like particles were refined and postprocessed, leading to a final reconstruction with an average resolution of 7.7 Å. In this final reconstruction, the density assigned to Ubp6 was not very well resolved, indicating heterogeneity of the particles in the s4 state. Therefore, the particles were further classified into four classes with a sphere mask around the Ubp6 density, using the focused classification described by Bohn et al. (54). Nearly half of the particles did not possess Ubp6 densities, whereas the rest of the particles showed a well-defined Ubp6 density.

From 3,304 micrographs, 5,680 sc26S and 40,709 dc26S particles remained after automated localization and 2D classification. The sc26S and dc26S particles were reconstructed separately. The resulting angles for dc26S were used for the generation of pseudo-sc26S particles, without prior polishing and dataset reduction. The resulting pseudo-sc26S and sc26S particles were merged and classified. Of six classes, two classes contained s3-like particles (67,522 particles). The particles were refined and postprocessed, leading to a final reconstruction with an average resolution of 7.8 Å.

The micrographs acquired with the Falcon II and Falcon III cameras were separately subjected to automated proteasome localization and 2D classification of the identified particles. For the Falcon II data, 22,621 sc26S and 35,254 dc26S particles were used for further analysis, and for the Falcon III data, 10,200 sc26S and 14,180 dc26S particles were used for further analysis. The sc26S and dc26S particles were reconstructed separately. The resulting angles for dc26S were used for the generation of pseudo-sc26S particles, without prior polishing and dataset reduction. The sc26S and pseudo-sc26S particles from both cameras were combined and classified into six classes. The class averages could not be clearly assigned to any state.

The precise voxel size of the single-particle reconstruction was determined by a cross-correlation analysis using UCSF Chimera (53), maximizing the fit of the 2.3-Å resolution yeast CP crystal structure [PDB ID code 5cz4 (31)] for different voxel sizes. This atomic model of the yeast 20S proteasome was also used as an initial model for the CP of state s1. The opening of the gate in s4 was modeled based on PDB ID code 1z7q (32). The initial model for the RP of the human 26S proteasome was built through comparative modeling using Modeler (55). The RP of the human 26S proteasome [PDB ID code 5l4g (22)] was used as a template, and the Modeler-implemented alignment scheme was used for sequence alignment. For Rpn1 and Rpn2, the X-ray structure of the isolated yeast Rpn2 [PDB ID code 4ady (56)] was used as an additional template. Missing segments in the template or sequence regions with a sequence identity lower than 30% were predicted following the same strategy as described by Schweitzer et al. (22). The CP of the homology model was fitted into the density using UCSF Chimera and then fitted into the density using MDFF (27), which employs molecular dynamics to fit initial models into a density in real space, and thus permits protein flexibility while maintaining realistic protein conformations. MDFF runs were prepared with QwikMD (44) and analyzed using the MDFF graphical user interface of VMD (45). We used NAMD (46) with the correction map-corrected CHARMM22 force field (57) for conducting MDFF. During MDFF runs, restraints to preserve the secondary structure, chirality, and cis-peptide bonds were applied to avoid overfitting. We used cascade MDFF starting with a density low-pass-filtered to a resolution of 8 Å and a grid coupling of 0.3, followed by runs with the 4.1-Å resolution map. The coupling of the model to the density was weighted according to the local resolution of the density. The quality of fit of the secondary structure elements was checked through local cross-correlation calculation (58) implemented in VMD. Next, we refined the structure according to the density in real space by combining MDFF with de novo structure prediction as well as Monte Carlo-based backbone and side-chain rotamer search algorithms in an iterative manner following the strategy as described by Schweitzer et al. (22). For further refinement of the atomic coordinates in reciprocal space, the EM map and the Fourier shell correlation were converted into an MTZ file using the ban_mrc_to_mtz.py script (59). The atomic coordinates were then optimized against the Fourier coefficients using maximum-likelihood refinement in PHENIX (60). Differences between the refined model and the map were analyzed in Coot (61), and the model was adapted accordingly by the tools offered in Coot, followed by a second step of refinement in PHENIX, yielding the final model.

The initial structure for the yeast s1 state was the merged structure of PDB ID code 5cz4 (31) for the core and the homology model for the RP based on the human structure [PDB ID code 5l4g (22)]. The refined final structure of s1 was used to initiate an MDFF run into the density for the s2 state. The final refined structure of the s2 state was fitted through MDFF into the density of the s3 state, and the final s3 state structure was then fitted into the s4 density.

To determine the orientation of Rpn13, we performed an exhaustive 6D correlation scan, using the EM map of the best resolved class of a classification with a spherical mask on the RP (Fig. S2A), starting from the position on the yeast model of the s2 state (PDB ID code 4cr3) (21). Z-scores were calculated based on the cross-correlation values between the refined EM map and a density generated from the fitted atomic structure, as previously demonstrated (30) (Fig. S2 C–F). The angle and shifts found by the initial scan were refined in UCSF Chimera. With these refined alignment parameters, the scan was repeated (Fig. S2 E and F). To support the result, we analyzed the cross-linking MS data of the 26S proteasome, in which two confident linkages between Rpn2 and Rpn13 were identified (Table S2).