Soluble human ACE (sACE), which includes the entire extracellular region of ACE, was expressed in HEK293F cells and purified from culture supernatant using an ion exchanger (Source Q) followed by size exclusion chromatography (Superdex S200). Buffer conditions suitable for single particle cryo-EM analysis were identified and optimized using differential scanning fluorometry (Figure 1—figure supplement 1, Supplementary file 1). Purified sACE was then vitrified using a Vitrobot, and a dataset of ~3600 micrographs was collected on a 300 keV Titan Krios equipped with a Gatan K3 camera at the University of Chicago Advanced Electron Microscopy Facility (Supplementary file 2). The data was processed in cryoSPARC v4.4.0 (Figure 1—figure supplement 2). After filtering ~1.1 million picked particles to ~378 K using 2D and 3D classification, three classes were identified (Figure 1, Figure 1—figure supplement 2). Of these, ~268 K particles mapped to a dominant 2-domain class reconstructed as an sACE-N dimer, while ~21 K particles mapped to a minor 3-domain class containing two sACE-N and only one sACE-C, reminiscent of the recently published sACE cryo-EM structure (Lubbe et al., 2022). The remaining 89 K particles mapped to a full-length sACE dimer containing two sACE-N and two sACE-C, which was refined to 3.65 Å resolution (Figure 1C, Figure 1—figure supplements 2 and 3, Figure 1—video 1, Supplementary file 2).

The partial denaturation of macromolecules during vitrification has been attributed mainly to repeated exposure of the macromolecules to the air-water interface (Glaeser, 2018; Noble et al., 2018). The sACE-C domain was shown to be less stable than the sACE-N domain (Voronov et al., 2002), so we hypothesized that repeated exposure to the air-water interface could cause preferential denaturation of sACE-C. We have previously shown that preferential denaturation of human PreP occurred in one of its two homologous domains and that faster vitrification using Chameleon significantly prevented such preferential denaturation (Liang et al., 2022). Therefore, we collaborated with the National Center for CryoEM Access and Training (NCCAT) to prepare sACE grids using Chameleon (Darrow et al., 2019). Chameleon preparation reduced the time the sample spends on the grid before freezing by more than 10-fold. We collected a new dataset of ~18,000 micrographs, and our 2D classification revealed a higher percentage of particles for four-domain classes than those prepared using the Vitrobot (Figure 1B). However, our 3D classification still showed the existence of structures of the sACE-N only dimer and those containing two sACE-N and one sACE-C (Figure 1—figure supplement 4). This indicates that the reduction in vitrification time using Chameleon reduced but did not prevent preferential denaturation.

From these data, we obtained a map for a consensus ACE structure that achieved a resolution of 2.8 Å. However, while the core region showed excellent map quality, the map quality for the dynamic regions corresponding to the open-close transition of ACE catalytic domains was poor, suggesting multiple discrete conformational states exist. We thus performed 3D classification and reconstructed three distinct conformational states of full-length apo-ACE at resolutions of 2.99 Å, 3.05 Å, and 3.15 Å (Figure 1C, Figure 1—figure supplements 4 and 5, Figure 1—videos 2–4, Supplementary file 2). We refer to these ACE dimer structures by their resolutions, namely sACE-2.99, sACE-3.05, and sACE-3.15, along with the structure from data obtained using the Vitrobot, sACE-3.65. Given the rich structural information from more than 50 reported crystal structures of ACE and one recently reported cryo-EM study of ACE, below we focus on the new insights that our sACE dimer structures offer (Lubbe et al., 2020; Acharya et al., 2024; Masuyer et al., 2012; Lubbe et al., 2022; Corradi et al., 2006; Natesh et al., 2003).

Soluble ACE has two homologous catalytic domains: sACE-N (amino acids 1–615) and sACE-C (amino acids 616–1202). Based on the comparison of various catalytic domains of our ACE dimer structures, we subdivide each catalytic domain of sACE into three subdomains: D1a, D1b, and D2 for sACE-N, and D3a, D3b, and D4 for sACE-C (Figure 1A). These three subdomains in either sACE-N or sACE-C form a catalytic chamber. The catalytic zinc ion is coordinated by the canonical HEXXH motif residing in the D1b and D3b subdomains (Figure 1A). Four sACE dimer structures manifest as pseudo-symmetric C2 homodimers across the dimer interface. The dominant symmetry-breaking features are the degrees of displacement between D1 and D2 or between D3 and D4, leading to changes in solvent accessibility to the catalytic pockets.

Based on the openness of the catalytic chamber, defined as the distance between the D1/3 a and D2/4 helices that gate access, we classified sACE-N into three conformational states: open (O), intermediate open (I), and closed (C) (Figure 2A); sACE-C is significantly less open than sACE-N in our structures and is only classified into intermediate (I) and closed (C) states (Figure 2B). The four sACE dimer structures differ mostly by the degrees of openness of their four catalytic domains (Figure 2C and D). Our sACE-2.99 and sACE-3.65 structures are globally similar, both contain sACE-N domains in the open and intermediate states and sACE-C domains in the closed and intermediate states. In sACE-3.05, both sACE-N domains adopt an open state, while the sACE-C domains adopt an intermediate and closed state. In the sACE-3.15 structure, one subunit has both sACE-N and sACE-C adopting a closed conformation, while the other subunit has an open sACE-N domain and an intermediate sACE-C domain.

Our four cryo-EM structures of sACE provide the structural basis for sACE dimerization. sACE dimerization is mediated by interfaces between D2 subdomains and D4 subdomains of sACE monomers. The D2 and D4 subdomains generally have much lower B-factors and higher local resolution estimates than the D1 and D3 subdomains (Figure 1—figure supplement 3, Figure 1—figure supplement 5, Figure 2—figure supplement 1). The interaction of the D4-D4 subdomains is analogous to the D2-D2 interaction, yielding an RMSD of ~1 Ų when aligned (Figure 3A). The primary source of variation is a general rigid body shift away from the interaction interface for non-interacting regions in the D4-D4 region compared to the D2-D2 region. The observed D2-D2 interface is consistent with one of three possible contacts between sACE-N in the non-crystallographic symmetry mates from previous crystallographic analysis of sACE-N and the previously published cryo-EM structure (Figure 3—figure supplement 1). In total, the interface between sACE monomers buries a surface area of approximately 1900 Ų. Although sACE-N and sACE-C are structurally analogous, their respective interfaces (sACE-N/N and sACE-C/C) are not equivalent. The sACE-N/N interface is larger and stronger, burying a surface area of approximately 1,150 Ų, compared to a buried surface area of approximately 750 Ų for the sACE-C/C interface.

The interface between sACE-N is formed by a hydrophobic core, anchored by a Y465 π-stacking interaction, and flanked by salt bridges formed by R453 and R458 as well as E212 and E219 along with a collection of polar and van der Waals contacts, which is consistent with the previous mutational studies of Y465 (Figure 3B; Danilov et al., 2011). The dimer interface between two sACE-N domains is completed by glycan-glycan interactions between the glycans attached to N82, which is consistent with the role of glycosylation in sACE dimerization (Figure 3C; Kost et al., 2000). However, the central residues that form the hydrophobic core of the interface between sACE-N are not present in the interface between sACE-C. Instead, the sACE-C interface is formed by a symmetric salt bridge between K1067 and E1060, side chain-backbone interactions, and van der Waals contacts (Figure 3D).

The comparison of our dimeric sACE cryo-EM structures reveals the conformational dynamics of sACE catalytic domains. The four sACE dimer structures represent different combinations of open, intermediate, and closed states of sACE catalytic domains (Figure 2). By aligning each sACE-N and sACE-C region in our structures to the D2/4 domain, it becomes clear that our structures represent a continuous gradient of structural heterogeneity, ranging from the most closed state to the most open state (Figure 2—figure supplement 2).

We performed a series of all-atom molecular dynamics (MD) simulations to explore the conformational dynamics of sACE. A total of eight apo-sACE simulations were conducted, comprising four simulations with non-glycosylated sACE (>1.8 μs each) and four with sACE where 12 N-linked glycosylation sites that contained observable extra density were modeled with complex glycans (>2 μs each; Figure 4—videos 1–2). Glycan orientations were randomized prior to each simulation to avoid bias. We did not observe a substantial difference in the global conformational dynamics between our sets of simulations with and without glycans. The domain motions of these simulations are summarized in Supplementary file 3.

RMSD analysis reveals that the dominant source of conformational change in our simulations is the open-close transition between the D1/3 and D2/4 domains (Figure 4—figure supplement 1). The D2/D2 and D4/D4 regions are the most stable, with the bulk of conformational changes localized to the D1/D3 domain. Within the D1/3 domain, we observed that the D1/3 a motion was not always correlated with D1/3b or D2/D4, meaning that at a given time, D1/3 a could move along with D1/3b, or with D2/D4, or independently. Consistent with our structural analysis, our MD simulations also indicate that sACE-N displays a much wider range of opening geometries than sACE-C.

Most of our MD simulations rapidly closed and remained closed for the duration of the simulation, preventing an effective comparison of the open vs closed state dynamics. However, we were able to garner significant insight into the mechanism of open-close transition by integrating our MD simulations and cryo-EM structural analysis. Initially, we sought to understand the stable and mobile regions of sACE by calculating the Cα displacement for each residue among the four difference cryo-EM structures. As shown in Figure 4A, the D2 domain was found to be the most stable region of sACE while the D1 domain was the most mobile, highlighting the impact of the sACE-N open-close transition. sACE-C was, globally, very stable outside of surface-exposed loops and the tip of the D3a subdomain. This analysis predicted the most mobile regions of sACE to be the ends of the D1a subdomain and a loop region with the D1b that closely associates with the tip of the D1a subdomain.

While the structural perturbations predicted by comparing our cryo-EM structures correlate well with the global RMSD analysis of our MD simulations (Figure 4—figure supplement 1), we performed further analysis to understand the finer details of sACE conformational dynamics. To best compare the cryo-EM structures with our MD simulations, we compared the calculated alpha carbon displacement of our cryo-EM structures to the root mean square fluctuation (RMSF) values for each residue in our MD simulations (Figure 4). We observed consistent, global agreement between our cryo-EM data and MD simulations, both with and without glycans present. This consistency is particularly noticeable in several key mobile regions (denoted 1–4 in Figure 4) that demonstrate high degrees of Cα displacement from our cryo-EM data along with correlating higher RMSF during subsets of our MD simulations (Figure 4A, B and D). The RMSF variation between stable and mobile regions in our MD simulations is dampened, relative to the cryo-EM data when the entire trajectories were used (Figure 4A, C and E). This is due to the open state regions closing relatively early in the simulation and remaining closed for the duration of the simulation. We then performed principal component analysis (PCA; Figure 4—figure supplement 2) and cluster analysis (Figure 4—figure supplement 3) to compare our MD simulation data with our cryo-EM data. Both methods of analysis indicate that the conformational space sampled by our cryo-EM structures represents a fraction of the conformational space sampled by our MD simulations.

The dynamics observed in our MD simulations of sACE are highly complex, arising from a combination of both interdomain and intradomain motions. To summarize these simulations in simpler terms, individual sACE domains transition from a more open state to a closed state without reopening, while sACE-N/N and sACE-C/C primarily move relative to each other as rigid bodies. As outlined previously, PCA was employed to reduce the complexity of these motions and to identify dominant trends in conformational changes within the dataset. Visualizing the conformational changes captured by the principal component vectors in the MD simulations (Figure 4—videos 3–4) revealed that the majority of conformational variability within the MD data is driven by substantial interdomain motion occurring between sACE-N/N and sACE-C/C. Consistent interdomain motion was observed along all principal component vectors, regardless of the presence or absence of intradomain motion. However, no principal components were identified that exclusively described significant intradomain motion without associated interdomain motion. The PCA further revealed that glycosylation significantly enhances the degree of interdomain motion in sACE, as evidenced by the glycosylated simulation (Figure 4—video 4) compared to the non-glycosylated simulation (Figure 4—video 3). This suggests that glycosylated sACE exhibits greater global dynamics (Figure 4, Figure 4—figure supplement 4). Notably, principal components 5 and 6 from the glycosylated MD simulations (Figure 4—video 4) highlight a striking difference, where sACE-N/N and sACE-C/C appear to move independently and exhibit a decoupled interaction. We postulate that this inter-domain motion may be the reason we observe preferential denaturation of sACE-C during cryo-EM grid preparation. The relatively harsh conditions that the protein is exposed to during the grid-making process may damage specific conformations of particles and subsequently restrict the amount of information that could be obtained solely from cryo-EM analysis. This highlights the importance of integrating multiple approaches to investigate protein dynamics.

The MD simulations of glycosylated sACE provide additional structural insights into the structure and functions of N-linked glycosylation of sACE. Glycosylation is shown to affect the folding, secretion, and stability of ACE (Kost et al., 2003; O’Neill et al., 2008; Sadhukhan and Sen, 1996). Our MD simulation reveals that the dynamics of 12 N-linked glycans cover a large surface area of sACE, shielding it from environmental insults (Figure 4—figure supplement 5). Consistent with observations from our Coulomb potential density maps and previously published data, we observed a substantial and persistent inter-subunit interaction between glycans from N82 (Figure 3C). From RMSF analysis, glycosylation was found to increase the overall dynamicism of sACE (Figure 4—figure supplement 4). Furthermore, while there are N-linked glycans near the catalytic chamber, particularly those at the tip of D1a and D3a, they were highly dynamic and did not occlude access to the catalytic chamber. This is consistent with the observation that deglycosylation only modestly increases the enzymatic activity of ACE (Baudin et al., 1997; Orth et al., 1998; Yu et al., 1997).

The comparison of sACE-N in its open and closed states reveals the structural basis for the open-to-closed transition of this domain. Our structures show that when the bottom of the D1a subdomain (region of interest 2 in Figure 4) moves away from the D2 subdomain, resulting in the open state, the top of the D1a subdomain (region of interest 1 in Figure 4) moves closer to the D2 subdomain, as if the helix were balanced upon a fulcrum (Figure 5A and B). Detailed structural analysis indicates that the fulcrum is primarily formed by the salt bridge between two highly conserved residues: E66 in the D1a subdomain and R108 in the D2 subdomain. To stabilize the open state, additional salt bridges are formed, with D189 in the D2 subdomain pairing with K73 in all cases and occasionally with R96 in the D1a subdomain (Figure 5A). The loss of this salt bridge in the closed state is compensated for by the formation of a network of van der Waals interactions between the D1a and D2 subdomains that stabilize the closed state. The closed state of sACE-C is also stabilized by a network of van der Waals contacts between the D3a and D4 subdomains. The transition between open and closed states is likely influenced by the binding of substrates and reaction products during the ACE catalytic cycle. Although a similar fulcrum mechanism is not evident in the sACE-C domain, the connection between the D3a and D2 subdomains would act as a lever to pivot sACE-C from the closed to the open state, and we speculate that this may form the structural basis for allosteric communication between sACE-N and sACE-C.

On average, sACE-N is more open than sACE-C, likely due to multiple factors (Figure 2D). In sACE-N, the D1a subdomain is at the N-terminal end of sACE, allowing for a greater range of motion (Figure 5B), while the fulcrum and stabilization mechanisms can maintain sACE-N in a more widely open state (Figure 5A). Conversely, in sACE-C, the motion of the D3a subdomain is constrained by its connection to the D2 subdomain (Figure 5C). Additionally, there are sequence differences at the interface between the three subdomains of sACE-N and sACE-C. In the most closed state—the B chain sACE-C domain of the sACE-3.05 structure—residues from all three subdomains, including I649, V724, V926, and V927, form a hydrophobic latch that stabilizes the closed state (Figure 5D). The hydrophobic latch in sACE-N is weaker because the residue corresponding to V724 has been replaced with a threonine (T124; Figure 5D). These structural insights are supported by our RMSF analysis of our MD simulations and alpha carbon displacement analysis of our cryo-EM structures, which reveal a correlation between mobile regions and the ‘fulcrum’ mechanism of opening for sACE-N (Figures 4 and 5A). Furthermore, our slowest closing MD simulation revealed that the K73-D189 interaction was maintained for the entire time that sACE-N adopted an open conformation, yet this interaction was lost when sACE-N transitioned to the closed state (Figure 5E).

We also performed three MD simulations on the non-glycosylated sACE monomer (>1 microsecond each) (Supplementary file 3). We observed the expected intra-subdomain open-closed motions in sACE-N and sACE-C that were also observed in the simulations of ACE dimer. We also observed that the motion of sACE-C relative to sACE-N was significantly greater in our monomer simulations compared to our dimer simulations. This suggests that dimerization places significant constraints upon the conformational dynamics of sACE. Previous work has defined the conformational dynamics of sACE from cryo-EM data where both sACE-C domains were not resolved (Liang et al., 2022). Given our observations from the monomeric and dimeric MD simulations, we endeavored to understand the conformational heterogeneity present within our cryo-EM data to better understand the impact of dimerization on sACE dynamics.

Cryo-EM structures are ensemble averages of many aligned particle images. Recently, there have been many approaches developed to understand the particle heterogeneity present within a cryo-EM structure as a proxy for molecular motion (Tang et al., 2023). We employed three methods of heterogeneity analysis: cryoDRGN, RECOVAR, and 3D Variability Analysis (3DVA; Punjani and Fleet, 2023; Gilles and Singer, 2024; Zhong et al., 2021). CryoDRGN employs neural networks to resolve structural heterogeneity in a nonlinear, multidimensional latent space, while RECOVAR and 3DVA both employ linear subspace methods. 3DVA finds the linear subspace by an alternating optimization scheme to capture image variations, while RECOVAR utilizes a regularized covariance estimator to identify subspaces that best capture the distribution of states.

We conducted this analysis and observed global agreement between the results of both of our datasets. As we observed substantial particle damage in our Vitrobot-prepared dataset, we thought it unwise to utilize that dataset for in-depth analysis and therefore, we focus on the larger and higher-resolution dataset collected at NCCAT below. For analysis, the entire particle set, prior to 3D classification and consisting of ~170 k particles, was utilized. The particle alignments from the same C1 consensus refinement were used as input for each of the three methods as current methods to investigate particle heterogeneity are highly dependent upon initial pose assignment of individual particles. Globally, all three methods agreed with our ensemble structures, in that the dominant source of structural variability within our particle population could be described by the open-closed transition of individual domains, and that sACE-N displayed more variability than sACE-C.

Initially, we performed 3D variability analysis (3DVA) in cryoSPARC to provide a direct comparison to the previously published analysis of the partial ACE cryo-EM structure (Punjani and Fleet, 2021; Figure 6A). Trajectories were reconstructed along the top 5 principal components of variance (Figure 6—video 1). Surprisingly, the trajectory describing the variance along the first principal component vector suggests a dominant interdomain motion, where the sACE-N/N region bends towards the sACE-C/C region, similar to what we observed for trajectories along the top principal component vectors in analysis of our MD simulations (Figure 4—videos 3–4), yet this variance is not reflected in our ensemble structures. The trajectories along the remaining principal component vectors comprise varying combinations of individual domains undergoing the open-close transition. Consistent with our previous observations, the range of presumptive motion in these trajectories is greater in sACE-N than sACE-C. Previous 3DVA performed on the sACE monomer structure revealed 4 types of potential motion, termed bending, pivoting, breathing, and jumping (Lubbe et al., 2022). The breathing motion matches the open-close transition that dominates most of our principal component vectors, while the bending motion appears to correlate with the variance described by our top principal component vector. We did not observe any potential motions reminiscent of the jumping or pivoting motions in our 3DVA. Furthermore, the magnitude of the bending motion in our 3DVA is substantially less than previously reported for the partial dimeric sACE structure, which likely suffered from substantial denaturation (Lubbe et al., 2022). Together, this highlights a significant difference in the conformational dynamics between our sACE dimer and previously reported cryo-EM structures. While there is disagreement between the 3DVA results for the two datasets, it should be noted that our MD simulations display a greater range of motion than either cryo-EM dataset offers, and the varieties of motion described previously are clearly observed within the PCA trajectories, especially when sACE is glycosylated (Figure 4—video 4). As mentioned above, this type of cryo-EM analysis is sensitive to artifacts introduced by particle damage and denaturation, underscoring the importance of preserving protein integrity via sample preparation and the use of Chameleon to reduce the exposure to air-water interface. This also suggests dimerization functions as a key constraint to the conformational dynamics of sACE.

We then moved on to analyze our data using cryoDRGN, which revealed that the data formed a single cluster across multiple dimensions (Figure 6, Figure 6—figure supplement 1). Trajectories generated along five principal component vectors of structural variance were dominated by the open-close transition, primarily in sACE-N (Figure 6—video 2). We observed little variance in sACE-C and, unlike 3DVA, cryoDRGN did not reveal any substantial interdomain motions. Faced with this discrepancy between the 3DVA and cryoDRGN results, we employed a third method, RECOVAR, to further analyze our data. Similar to cryoDRGN, RECOVAR revealed that our data formed a singular cluster across multiple dimensions of analysis. Volumes representative of k-means cluster reconstructions were similarly consistent with our cryoDRGN results. RECOVAR indicated that the primary source of structural variance could be explained by the open-close transition, primarily of the sACE-N domains. With this in mind, we repeated the RECOVAR analysis with a focus mask around the sACE-N/N region. We then estimated the conformational density by deconvoluting the multidimensional latent space with a two-dimensional principal component analysis. The resulting density plot revealed four clusters that corresponded to different conformations of the sACE-N/N region (Figure 6C). The most populous cluster corresponded to a state where both sACE-N subunits adopted an open conformation. The second largest cluster, and the smallest cluster, corresponded to a state where one sACE-N was open and the other was closed. The discrepancy in population size between these clusters is likely due to bias in the initial particle orientation, rather than a subunit-specific preference for the open state. As the O/C state and the C/O state are 180 degree rotations of each other, particle assignment to either cluster is likely influenced by the initial particle orientation of the complete dimer, and we currently lack the data to discern any allosteric implication to the orientation assignment. The final cluster corresponded to a state where both sACE-N subunits adopted the closed conformation. We then used RECOVAR to generate a trajectory through the latent space to predict a potential transition pathway from the closed-closed state to the open-open state (Figure 6—video 3). This trajectory revealed that one subunit would open, followed by the second, and raised the question of potential allostery or coordinated motion between the adjacent sACE-N domains.

To investigate coordination between the sACE-N domains, we performed 3D classification in cryoSPARC with a focus mask on a single sACE-N domain and requested two classes (Figure 6D). The particles were split roughly evenly between the two classes, one resulting in an open state and the other being closed. We then used these classes for another round of 3D classification with a focus mask on the opposite sACE-N domain and again asked for two classes. Again, these particles were split roughly evenly among the classes, leaving us with four similarly sized classes adopting the same conformations as the clusters revealed by RECOVAR.

Human 293T and 293 F cells were obtained from ATCC (CRL-3216) and Thermo Fisher (R79007), respectively. Cell lines were subjected to re-authentication on a regular basis to ensure no cross-contamination. All cells are routinely screened for the presence of mycoplasma using the ATCC Universal Mycoplasma Detection Kit (Catalogue # 30–1012 K). These cells were maintained in DMEM +10% fetal calf serum (FCS). The lentiviral plasmid for the expression of the extracellular domain of human ACE (sACE, aa 1–1235) that contains a 6-histidine tag at its C-terminus was made by Vectorbuilder and co-transfected with helper plasmids to human 293T cells for lentiviral production. Human 293 F cells were infected with the resulting virus and selected for stable sACE expression. 293 F that stably expressed sACE grew poorly in Free-style 293 media (ThermoFisher 12338018). Thus, for ACE production, 293 F cells that stably expressed sACE were first grown in DMEM +10% FCS at 37 °C and gradually shifted into Free-style 293 media by reducing FCS by first in Free-style 293 media +2% FCS and then to media without FCS. These cells were then shifted to and maintained in 30 °C for 3–5 days for ACE production to reduce proteolysis.

Culture supernatant was centrifuged at 3000x g for 5 min, diluted 10-fold with 25 mM Tris-HCl pH 8, filtered, and run over a Source Q column pre-equilibrated with 25 mM Tris-HCl pH 8. The column was washed to baseline with Tris-HCl pH 8 and bound protein was eluted with a linear gradient of NaCl to 1 M. sACE containing fractions were pooled and EDTA was added to 10 mM to strip the catalytic zinc ion. The sample was concentrated and run over a Superdex S200 SEC column pre-equilibrated with the desired cryo-EM buffer.

To optimize the grid making conditions, differential scanning fluorimetry was applied to screen 96 buffers conditions. DSF was performed with 3 mg/ml sACE and 10 X Sypro Orange in 20 µl buffers in a Thermo Fisher Step ONE RT-PCR thermocycler. Using the melting temperature (T_M) and slope as selective criteria, 25 mM citrate pH 5.5, 25 mM sodium phosphate pH 5.5, 150 mM NaCl, 10 mM EDTA was identified as the optimal condition for the highest T_M and optimal slope, which was used for all subsequent cryo-EM experiments.

Peak fractions were collected off the SEC column and used to immediately prepare grids, resulting in a protein concentration of ~4.5 mg/ml. Quantifoil R1.2/1.3 Cu200 grids were glow-discharged in air and grids were prepared by Vitrobot mark 3. 3.5 μl of sample was applied to the grid in the chamber with 100% humidity, 25 °C. Excess sample was blotted for 1 s, blot force 2, and grids were plunge frozen in liquid ethane. All images were acquired using a Titan Krios microscope (FEI) operated at 300 KeV with a Gatan K3 direct electron detector. A dataset of ~3600 micrographs was collected at The University of Chicago Advanced Electron Microscopy facility. The dataset was processed using cryoSPARC version 4.5.1 and 4.6.0 (Figure 1—figure supplement 2; Punjani et al., 2017). Briefly, a four-domain sACE dimer model was generated using sACE-N dimer (PDB 7Q4D) and sACE monomer (PDB 7Q3Y) and adjusting the distance between ACE-C domains to avoid steric clashes as a template for particle picking (Lubbe et al., 2022). After template picking, 1.1 million particles were extracted with a box size of 360 pixels. Several rounds of 2D classification were used to remove junk particles.~577 K particles were processed in successive rounds of ab initio reconstruction and heterogeneous refinement. sACE grids were also prepared using Chameleon and collected a dataset of ~19,000 micrographs at the National Center for Cryo-EM Access and Training. For those, Quantifoil Active R 1.6/0.9 Cu/Rh 300 was coated with gold using a Safematic CCU-010 Compact Coating Unit. Briefly, grids were placed nanowire side down on a glass slide, gold disks were sputtered with a 35 nm target thickness at 1.5x10^–2 Torr, and a process current of 30 mA. Grids were glow discharged with air at 12 mA for 80 s using the internal Chameleon glow discharger. The grids were plunged at 160–180 ms. All images were acquired using a Titan Krios microscope (FEI) operated at 300 KeV with a Gatan K3 direct electron detector (Gatan). Images were automatically acquired using Leginon (Suloway et al., 2005) using collection parameters as shown in Supplementary file 2. This dataset was processed in cryoSPARC version 4.5.1 and 4.6.0 (Figure 1—figure supplement 4). Briefly, templates were generated from our full-length ACE structure described above using vitrobot-prepared grids. Micrographs with a CTF fit resolution worse than 4 Å were removed from the dataset, leaving ~12,000 micrographs for template picking, which yielded 10.7 million particles. The particle set was trimmed to ~2.8 million based on quality metrics, particles were extracted with a box size of 256 pixels and run through 2 rounds of 2D classification, reducing the particle number to ~660,000. 2D classes were dominated by 4 domain views and yielded a much greater diversity of views, indicating that Chameleon preparation produced a dramatic improvement in particle quality. Initial 3D homogeneous refinement reached 3.6 Å, matching the previous final structure. Four maps from 3D classes that contain all four catalytic domains of ACE, one from particles derived from grids prepared by using vitrobot and three from those prepared by using Chameleon, were used for structure building using Chimera and Coot, and the structure was refined using Phenix real space refinement. Structural analysis was performed using PYMOL and Chimera.

All heterogeneity analysis was performed on the full final particle set, encompassing all of the NCCAT structures (Supplementary file 4). Initially, cryoSPARC 3DVA was performed on both the NCCAT- and UChicago-collected datasets, and results were found to be consistent, so the NCCAT-collected dataset was used for further analysis, given its higher particle number and quality. Input particle poses for each method of heterogeneity analysis were derived from the same C1 refinement job in cryoSPARC, which reached 2.88 Å. 3DVA was used to solve the top 5 modes of variance with a filter resolution of 6 Å. Particles were downsampled to a box size of 128 pixels in cryoDRGN 3.3.1 to speed up processing. The same downsampled particle set was used for both cryoDRGN and RECOVAR. CryoDRGN was run with default parameters and the analysis was performed on the top 5 principal components for direct comparison to cryoSPARC 3DVA. RECOVAR pipeline.py was initially run on the full sACE structure. Reconstructions representative of k-means cluster centers revealed the dominant source of structural variability to be the open-close transition in sACE-N. As a result, RECOVAR was run again with a focus mask on sACE-N/N and a zdim = 4. Conformational density was estimated with a two-dimensional PCA, and an index value of 3 was chosen for further analysis. The points in latent space representing the top 12 stable conformational states were estimated from the top 20% of particles and representative volumes were generated. The generated states populated four clusters, and the 12 states were simplified to 4 distinct conformations where sACE-N/N adopted the open-open (OO), open-closed (OC), closed-open (CO), and closed-closed (CC) states (Figure 6C). We then generated a trajectory through the latent space using the OO and CC representative volumes as endpoints (Figure 6—video 3).

The sACE-3.65 structure was used to initiate all MD simulations. For the glycosylated sACE simulations, N-linked glycans were built onto all 12 known N-linked glycosylation sites (residues 9, 25, 45, 82, 117, 289, 416, 480, 648, 666, 685, 731) with the complex glycan BGLC-BGLC-BMAN-AMAN-BGLC-ANE5-AMAN-BGLC-BGAL-ANE5-AFUC using CHARMM-GUI (Lee et al., 2016). Simulations were prepared in QwikMD or CHARMM-GUI and run at pH 7, 150 mM NaCl, 310 K, 1 atm with periodic boundary conditions in NAMD3.0 using the University of Chicago Beagle3 partition (Phillips et al., 2020). Explicit solvent was described with the TIP3P model (Jorgensen et al., 1983). Data was analyzed in VMD 1.9.4 (Humphrey et al., 1996) using a combination of built-in plugins and custom Python scripts. Analysis scripts are available as source files for the analysis of PCA (Supplementary file 5), cluster analysis (Supplementary file 6), C-alpha displacement (Supplementary file 7).