Structural basis for the ARF GAP activity and specificity of the C9orf72 complex

Mutation of C9ORF72 is the most common genetic cause of amyotrophic lateral sclerosis (ALS) and frontal temporal degeneration (FTD), which is attributed to both a gain and loss of function. C9orf72 forms a complex with SMCR8 and WDR41, which was reported to have GTPase activating protein activity toward ARF proteins, RAB8A, and RAB11A. We determined the cryo-EM structure of ARF1-GDP-BeF3- bound to C9orf72:SMCR8:WDR41. The SMCR8longin and C9orf72longin domains form the binding pocket for ARF1. One face of the C9orf72longin domain holds ARF1 in place, while the SMCR8longin positions the catalytic finger Arg147 in the ARF1 active site. Mutations in interfacial residues of ARF1 and C9orf72 reduced or eliminated GAP activity. RAB8A GAP required ~10-fold higher concentrations of the C9orf72 complex than for ARF1. These data support a specific function for the C9orf72 complex as an ARF GAP. The structure also provides a model for the active forms of the longin domain GAPs of FLCN and NPRL2 that regulate the Rag GTPases of the mTORC1 pathway.


Results
We first characterized the concentration of C9orf72:SMCR8: WDR41 needed to produce complete hydrolysis of GTP by ARF1. An ARF1 construct with the N-terminal myristoylated amphipathic helix removed was used in all experiments, as this region was previously found to be unnecessary for its competence as a GAP substrate 21 . The concentration of ARF1 was fixed at 30 µM, and then treated with 0, 0.3, 0.6, and 3-µM C9orf72 complex for 15 min at 37°C, and nucleotide products were monitored by high pressure liquid chromatogram (HPLC) analysis. The 0.3 µM C9orf72:SMCR8:WDR41 sample, equivalent to a 1/100 molar ratio relative to ARF1, was sufficient to completely hydrolyze all of the GTP-bound to ARF1 (Supplementary Fig. 1a).
To probe the structural basis of the catalytic mechanism, we attempted to assemble the GAP-ARF1 complex by mixing the GTP-locked mutant ARF1 Q71L with the C9orf72:SMCR8:WDR41 protein complex. A stable complex was observed by gel-filtration chromatography ( Supplementary Fig. 2a, b), but not by cryo-EM ( Supplementary Fig. 2c). We inferred that the complex dissociated during vitrification. Following a precedent in which it was found necessary to use a fusion construct to trap the structure of the ASAP3:ARF6 GAP complex 30 , we fused the C terminus of ARF1 to the flexible N-terminus of SMCR8, and coexpressed the fusion with C9orf72 and WDR41 in HEK293-GnTi cells as a chimeric C9orf72:ARF1-SMCR8:WDR41 complex (Fig. 1a). The N-terminal 50 residues of SMCR8 are disordered 21 , effectively serving as a flexible linker between ARF1 and SMCR8 of up to 150 Å in length, which is similar to the longest axis of the C9orf72:SMCR8:WDR41 complex. The size exclusion profile demonstrated a similar yield and hydrodynamic behavior to the wild-type complex (Fig. 1b). HPLC analysis was used to assay the identity of the nucleotide bound in ARF1. Recombinant ARF1 protein as isolated from cells was 45-55% GTP-loaded, whereas the fusion version was only GDP-bound (Fig. 1c). This suggested to us that the fusion construct was catalytically active.
In order to capture a stable complex, we incubated the C9orf72: ARF1-SMCR8:WDR41 complex with both ground-or transitionstate analogs (BeF 3 − and AlF x − ) and then tested if they could still promote the GTP hydrolysis for the purified ARF1. In the presence of BeF 3 − , but not AlF x − , GAP activity was inhibited. Thus, we generated a nucleotide hydrolysis ground-state analog 31 by incubating the GDP-bound protein complex with BeF 3 − overnight, and visualized it by cryo-EM. A preliminary reconstruction of C9orf72:ARF1-SMCR8:WDR41 was obtained from a dataset collected on a 200-kV Talos Arctica. This reconstruction showed extra density corresponding to ARF1 localized atop the two longin domains ( Supplementary Fig. 3), and coordinates for the GTP-bound ARF1 model (PDB 1O3Y) 32 fit the density. However, some regions were poorly resolved, and additional datasets were collected on a 300-kV Titan Krios microscope . We were able to obtain a 3.9-Å cryo-EM reconstruction from an untilted dataset combined with 30°and 40°stage tilted datasets. It was possible to visualize the locations of the bound nucleotide and larger side-chains in this reconstruction, but smaller side-chains were not resolvable, nor were conformational details of larger side-chains.
The reconstruction showed that ARF1 sits in a 25 Å-wide concave pocket generated by the SMCR8 longin and C9orf72 longin domains. The ARF1 switch I and switch II regions, which are conformationally sensitive to the nucleotide state, directly contact the parallel helices of the SMCR8 longin -C9orf72 longin dimer (Fig. 2a, b). The bowl formed by the two longin domains can be compared to the bowl of a spoon, and the parallel N-terminal helices of SMCR8 to the spoon handle. Relative to C9orf72, SMCR8 longin has the more extensive interface with ARF1 (Fig. 2b). ARF1 is in its GTP-bound conformation, with respect to all three of the regions whose conformation depends on the nucleotide state: the switch I, switch II, and interswitch regions (Fig. 2c).
SMCR8 Arg147 points directly toward the BeF 3 − , as predicted by the arginine finger model (Fig. 3a, b, left). The geometry is similar to that seen in the only other structure of a catalytically competent ARF GAP complex, that of ASAP3 and ARF6-GDP-AlF x (PDB: 3LVQ) 30 (Fig. 3a, b, middle). In the ASAP3:ARF6structure, only one face of ARF6 contacts the GAP. In comparison, the C9orf72-SMCR8 longin spoon bowl cups ARF1 more extensively. ARF1 switch II sits on the parallel helices of the two longin domains and the loop of SMCR8 longin , while switch I binds only to SMCR8 longin . The more extensive cupping across the Switch I/II regions is seen in other GAP complexes, however, such as the TBC1D20:RAB1B-GDP-BeF 3 − complex (PDB: 4HLQ) 33 (Fig. 3a, b, right). ARF1 residues at the C9orf72 and SMCR8 interface were mutated and tested for GAP activity as monitored by HPLC (Fig. 4a, b). Mutation of residues (I49H, G50R) located at the switch I region largely reduced the activity compared to the ones at the switch II region (K73W), which is in consistent with the more extensive interface of SMCR8. We mutated the C9orf72 interface residues Trp33 and Asp34 in the double mutant W33G/ D34K (Fig. 4c). C9orf72 W33G/D34K :SMCR8:WDR41 complex was found to almost completely abolish GAP activity toward ARF1. We confirmed that C9orf72:SMCR8 was active toward RAB8A, albeit at much higher concentrations of the GAP complex than for ARF1. GAP activity with respect to RAB8A requires a 1/10 molar ratio of C9orf72:SMCR8:WDR41 to hydrolyze the RAB8A proteins, at least approximately tenfold more than for ARF1 ( Supplementary Fig. 1b). The I71R mutation of Ile71 on the C9orf72 longin helix decreases the activity toward ARF1, whilst essentially abolishing activity toward RAB8A (Fig. 4).

Discussion
C9orf72-SMCR8 is one of a class of dimeric longin-DENN domain complexes that also include the FLCN-FNIP2 complex. Catalytic Arg residues of GATOR1 subunit NPRL2 34 , FLCN 22,23 , and SMCR8 20,21 have been proposed to mediate GAP activity by an Arg finger mechanism 35 . We have now trapped and visualized the ARF1 GAP complex of C9orf72:SMCR8, confirming that the proposed Arg147 residue of SMCR8 serves as a catalytic Arg finger in the expected geometry. Thus far it has not been possible to trap GATOR1 or FLCN-FNIP2 in catalytically competent GAP complexes with their substrates RagA and RagC, owing to the transient nature of the complexes and competition with more stable inhibitory states. This has led to questions about whether the Rags might not in fact be regulated by the Arg fingers of FLCN and NPRL2, and might instead be more similar to the Gα subunit of heterotrimeric GTPases 24 . The structure provided here provides a basis for more precise modeling of the GATOR1 and FLCN GAP complexes in the active conformation (Supplementary Fig. 7). These models show that engagement of the Arg fingers of FLCN and NPRL2 in the FLCN-FNIP2 and GATOR1 complexes is sterically allowed and consistent with the mode of binding of SMCR8 to ARF1.
The great majority of the >30 known ARF GAP proteins contain a conserved ARF GAP domain, first described for ArfGAP1 36 . Just a handful of ARF GAPs function without a canonical ARF GAP domain, including SEC23, the GAP for the ARF-related SAR1 GTPase, and the ELMO domain proteins ELMOD1-3 36 . The data presented above define the C9orf72: SMCR8 longin-DENN complex as a new addition to the small set of protein folds possessing ARF GAP activity. We previously found that C9orf72:SMCR8 was active against ARF1, ARF5, and ARF6, although it had no activity against ARL8A and B 21 . The epitope immediately surrounding the GTP-binding site of ARFs 1-6 is highly conserved, and GAPs and GEFs are therefore typically active across the family 37 . ARF6 regulates axon extension 38 , and dominant-negative ARF6 expression rescues the axonal actin dynamics phenotype in C9ORF72 ALS patient cells 17 . While ARF6 is the best candidate for a C9orf72 GAP substrate relevant to neurodegeneration, the structural features of the ARF1 complex presented here are expected to be essentially identical between the two.
Our data strongly support that C9orf72:SMCR8 is a GAP for ARF GTPases. Given the report that C9orf72:SMCR8 is also a GAP for RAB8A and RAB11A 20 , the question as to what are the physiological substrates of C9orf72:SMCR8 has become important in C9ORF72, ALS, and FTD research 39,40 . The data presented here affirm C9orf72:SMCR8 as a bona fide ARF GAP by presenting a well-ordered cryo-EM structure, which makes verifiable predictions of the activity of point mutants of both ARF1 and C9orf72. The generally much lower activity measured for RAB8A leaves the functional significance of the RAB GAP activity uncertain. It is possible that the presence of lipids or as-yet unidentified protein factors or post-translational modifications could increase the efficiency of RAB8A as a C9orf72:SMCR8 GAP substrate.
Another difference between the structures presented here and previously 21 is that our complex, expressed in HEK293 cells was monomeric, while material expressed in insect Sf9 cells was dimeric 20 . We noticed a small population of C9orf72:SMCR8: WDR41 dimeric complex in our cryo-EM sample, and the dimeric complex was more populated when we cross-linked it at high concentrations of protein complex. This suggests that there is monomer-dimer equilibrium in the solution. Our observation of a catalytically competent monomeric complex does show that the monomer is the minimal entity capable of GAP activity.
The C9orf72 complex is reversibly localized to lysosomes in amino acid starvation through its WDR41 subunit 41 . WDR41 in turn binds to the lysosomal cationic amino acid transporter PQLC2 42 . The PQLC2 binding site on WDR41 was recently identified 43 , and is near the distal tip of the overall complex with respect to the ARF binding site. It is not apparent, based on existing structural data, how C9orf72:SMCR8:WDR41 could be localized to the lysosomal membrane to carry out GAP activity in cis on a small GTPase localized to the same membrane. There are few reports of ARFs localized to lysosomes, and thus would not be surprising that the complex is not designed to act in cis while lysosomally localized. This leaves two models for lysosomal regulation of ARF GAP activity. In the sequestration model, C9orf72 is active as a GAP in amino acid replete cells, but inactivated by lysosomal sequestration in starved cells. Alternatively, in the trans-activation model, the ARF GAP active site could project away from the lysosome to act on a pool of ARF localized to nearby membranes or cytoskeletal elements under starvation conditions. The working hypothesis suggested by the data above is that C9orf72 haploinsufficiency in ALS and FTD results in an excess of ARF GTP in one or more localized pools within the cell. These might be in the vicinity of lysosomes and/or within axons. The locations and natures of these pools are major open questions whose answers call for further study. Given the numerous roles of ARF proteins in membrane traffic 36 , it would not be surprising if excess ARF GTP led to multiple defects in signaling, membrane traffic, and autophagy, which are correlates of C9ORF72 haploinsufficiency in disease 6,7 . This idea does suggest a potential therapeutic approach for ALS and FTD. In principle, ARF GTP levels could be rebalanced by inhibiting the ARF GEFs responsible for the GTP loading of the relevant local pools of ARF GTP , if the appropriate ARF GEF(s) were identified. This family of ARF activators comprises at least 15 members 36 , which are the targets of the relatively unspecific natural product brefeldin A. A number of highly specific ARF GEF inhibitors have been identified, and strategies for their screening are well worked out 44 , lending support to the plausibility of such an approach.

Methods
Protein expression and purification. The gene encoding ARF1 (residues E17-K181) was amplified and inserted into a pCAG vector encoding an N-terminal STREP-STREP-FLAG tagged SMCR8 using KpnI restriction sites (Supplementary  Table 1), resulting in a STREP-STREP-FLAG-ARF1-SMCR8 construct. The pCAG vector encoding an N-terminal glutathione S-transferase (GST) followed by a TEV cleavage site was used for expression of C9orf72. WDR41 was cloned into a pCAG vector without a tag. The complex was expressed in HEK293-GnTi cells. The gene encoding RAB8A (residues M1-S181) was subcloned into pCAG resulting in a STREP-STREP-FLAG-RAB8A. All the ARF1 mutants were created by site-directed mutagenesis using KAPA HiFi HotStart ReadyMix. The C9orf72 mutants were generated by two-step PCR and ligated into the pCAG-GST vector. HEK293-GnTi cells adapted for suspension were grown in Freestyle media supplemented with 1% FBS and 1% antibiotic-antimycotic at 37°C, 80% humidity, 5% CO 2 , and shaking at 140 rpm. Once the cultures reached 1.5-2 million cells mL −1 in the desired volume, they were transfected as follows. For a 1-L transfection, 3-mL PEI (1 mg ml −1 , pH 7.4, Polysciences) was added to 50-mL hybridoma media (Invitrogen) and 1 mg of total DNA (isolated from transformed E. coli XL10-gold) in another 50-mL hybridoma media. Each 1 mg of transfection DNA contained an equal mass ratio of the three C9orf72 complex subunit expression plasmids. PEI was added to the DNA, mixed, and incubated for 15 min at room temperature. One hundred milliliters of the transfection mix were then added to each 1-L culture. Cells were harvested after 3 days. Cells were lysed by gentle rocking in lysis buffer containing 50-mM HEPES, pH 7.4, 200-mM NaCl, 2-mM MgCl 2 , 1% (vol/vol) Triton X-100, 0.5-mM TCEP, protease inhibitors (AEBSF, Leupeptin, and Benzamidine) and supplemented with phosphatase inhibitors (50-mM NaF and 10-mM beta-glycerophosphate) at 4°C. Lysates were clarified by centrifugation (15,000 g for 40 min at 4°C) and incubated with 5-mL glutathione Sepharose 4B (GE Healthcare) for 1.5 h at 4°C with gentle shaking. The glutathione Sepharose 4B matrix was applied to a gravity column, washed with 100-mL wash buffer (20-mM HEPES, pH 7.4, 200-mM NaCl, 2-mM MgCl 2 , and 0.5-mM TCEP), and purified complexes were eluted with 40-mL wash buffer containing 50 mM reduced glutathione. Eluted complexes were treated with TEV protease at 4°C overnight. TEV-treated complexes were purified to homogeneity by injection on Superose 6 Increased 10/300 GL (GE Healthcare) column that was pre-equilibrated in gelfiltration buffer (20-mM HEPES, pH 7.4, 200-mM NaCl, 2-mM MgCl 2 , and 0.5-mM TCEP). For long-term storage, fractions from the gel-filtration chromatography were snap-frozen in liquid nitrogen and kept at −80°C.
For expression of His 6 -tagged ARF1 (residues E17-K181) and all the ARF1 mutants, the plasmids were transformed into E. coli BL21 DE3 star cells and induced with 1-mM IPTG for 3 h. The cells were lysed in 50-mM Tris-HCl pH 8.0, 300-mM NaCl, 2-mM MgCl 2 , 10-mM imidazole, 0.5-mM TCEP, and 1-mM PMSF by ultrasonication. The lysate was centrifuged at 15,000 × g for 30 min. The supernatant was loaded into Ni-NTA resin and washed with 20 and 50-mM imidazole and eluted with 300-mM imidazole. The eluate was further purified on a Superdex 75 10/300 GL (GE Healthcare) column equilibrated in 20-mM HEPES, pH 7.4, 200-mM NaCl, 2-mM MgCl 2 , and 0.5-mM TCEP. STREP-STREP-FLAG tag RAB8A (residues M1-S181) was expressed in HEK293-GnTi cells and purified by Strep resin and eluted in 10-mM desthiobiotin buffer. The eluted protein was applied on Superdex 75 10/300 GL column. The nucleotide-binding state of each small GTP-binding protein was assessed by HPLC (see below). The purified RAB8A was GDP-bound and our previous data demonstrated that it had spontaneous guanine exchange activity 21 , therefore RAB8A was incubated with GTP at room temperature for 2 h and applied to a PD 10 column to remove the excess nucleotide. The resulting RAB8A sample had roughly 45% GTP-bound. ARF1 I49H/G50R and ARF1 I49H/G50R/K73W were also GDP-bound, therefore the GTP reloading protocol was also applied to these samples.
The proteins used in this study are shown in Supplementary Fig. 8 and the uncropped gels are presented in Supplementary Fig 9. Cryo-electron microscopy grid preparation and data imaging.    Fig. 4 Interfaces between ARF1 and C9orf72/SMCR8 and HPLC result of ARF1 WT/mutants with C9orf72 WT/mutants :SMCR8:WDR41. a Close-up views of the interfaces between switch I/switch II regions of ARF1 with C9orf72:SMCR8. b HPLC analysis of ARF1 WT/mutants incubated with C9orf72:SMCR8: WDR41. The top panels are the control reactions for each ARF1 proteins, either wild-type or mutants. The bottom ones are incubated with 1/100 molar ratio of C9orf72:SMCR8:WDR41. c ARF1 or RAB8A incubated with 1/100 or 1/10 molar ratio of C9orf72 WT/mutants :SMCR8:WDR41, respectively. The experiments in b, c are carried out at least three times, and one representative plot is shown.
For the untilted Titan Krios dataset, grids were prepared in the same manner and the complex was collected using a Titan Krios G2 electron microscope (FEI) operating at 300 kV with a Gatan Quantum energy filter (operated at 20-eV slit width) using a K3 summit direct electron detector (Gatan, Inc.) in super-resolution counting mode at magnification 81,000×, corresponding to a pixel size of 0.47 Å on the specimen level. In total, 10,188 movies were collected with defocus of −1.2 to −2.5 μm with a 3 × 3 image shift pattern prior to stage movement. Movies consisted of 50 frames, with a total dose of 50 e − /Å 2 and a total exposure time of 2.929 s.
The Titan Krios dataset with a stage tilt of 30°was obtained using 4-μL sample applied on freshly glow-discharged (PELCO easiGlow, 25 s in air at 25 mA and 0.4 mbar) UltrAufoil R1.2/1.3 grid (Ted Pella) and plunged into liquid ethane after 30s incubation and blotting for 4 s using a Vitrobot Mark IV (FEI) with blot force 0 and 100% humidity and temperature set to 5°C. Four thousand five hundred and eighty-five movies were collected using SerialEM at around −1.5-2.8-μm defocus with 1 × 3 image shift pattern. The movies were exposed to a total dose of 50 e − /Å 2 for 5.326 s at CDS mode.
The Titan Krios dataset with a stage tilt of 40°was collected using 4 μL sample applied on freshly glow-discharged (PELCO easiGlow, 25 s in air at 25 mA and 0.4 mbar) GF-1.2/1.3-3Au-45nm-50 grid (Electron Microscopy Sciences) and plunged into liquid ethane after 30-s incubation and blotting for 3 s using a Vitrobot Mark IV (FEI) with blot force 8 and 100% humidity and temperature set to 4°C. Six thousand one hundred and seventy-eight movies were collected using SerialEM at −1.5-2.8-μm defocus with 1 × 3 image shift pattern. The movies were exposed to a total dose of 50 e − /Å 2 for 5.326 s at CDS mode.
All of the data were acquired with SerialEM using custom macros for automated single particle data acquisition. Imaging parameters for the dataset are summarized in Supplementary Table 2.
Cryo-electron microscopy data processing. For the Talos Arctica dataset, drift and beam-induced motion correction was applied on the super-resolution movie stacks using MotionCor2 45 and using 7 × 5 patches and binned 2x by Fourier cropping to a calibrated pixel size of 1.14 Å/pixel. The defocus values were estimated by CTFFIND4 46 from summed imaged without dose weighting. After manual check to remove empty or crystalline ice micrographs, 251 micrographs were discarded and the remaining 2328 micrographs were subjected to reference-free particle picking using gautomatch-v0.53 (http://www.mrc-lmb.cam.ac.uk/kzhang/ Gautomatch/). About four million particles were extracted with binning 4x in Relion 3.1 47 and applied to iterations of three-dimensional (3D) classifications. The best class was selected and re-extracted with binning 1x. The particles at the edges were removed and imported into cryoSPARC v2 48 for further two-dimensional (2D) classification clean up and non-uniform refinement. The resulting 80,569 particles were refined to 4.84 Å and used as reference map for further Titan Krios dataset processing. The workflow is depicted in Supplementary Fig. 3.
For the untilted Krios dataset, drift and beam-induced motion correction was applied on the super-resolution movie stacks using MotionCor2 using 7 × 5 patches and binned 2x to a calibrated pixel size of 0.94 Å/pixel. The defocus values were estimated by CTFFIND4 from summed imaged without dose weighting. Other processing procedures were performed within Relion 3.1. Seven thousand three hundred and forty-seven micrographs with CtfMaxResolution values better than 5 were selected for further processing. Around six millions of particles were picked using gautomatch-v0.53 (http://www.mrc-lmb.cam.ac.uk/kzhang/Gautomatch/) and then extracted with binning 6x. The particles were subjected to 3D classification using a 45-Å low-pass filtered reference generated by the reconstruction collected in Talos Arctica microscope. 888,231 particles from the best class were selected and refined to 7.755 Å. The refined particles were further cleaned up with rounds of 2D and 3D classification. The extracted, unbinned particles was refined to 4.5 Å and applied to CTF refinement for beam-tilt and perparticle defocus determination as well as Bayesian polishing. The shiny particles were further refined to 3.7 Å. Map sharpening was performed with DeepEMhancer 49 using the "tightTarget" or "highRes" models. The workflow is depicted in Supplementary Fig. 4.
For the Krios dataset with a stage tilt of 30°, data processing followed a similar pipeline of 0°tilt dataset for processing except that first run a full-micrographbased CTF estimation, then pick particles using gautomatch and then subjected the particle coordinates to rerun local CTF estimation on a per-particle based in gctf 50 ( Supplementary Fig. 4). For the Krios dataset with a stage tilt of 40°, the data were imported into cryoSPARC v2 for patch motion correction and patch CTF estimation. Particle picking was done using 2D class averages from 0°tilt as reference for Autopick in Relion 3.1 and subjected to several rounds of 3D classification. The best classes were imported into cryoSPARC v2 for further rounds of 2D classification. The clean particles were applied to both global and local CTF refinement for then imported into Relion 3.1 to merge with other two datasets ( Supplementary Fig. 4).
Atomic model building and refinement. For the C9orf72:ARF1-SMCR8: WDR41 model, one copy of the coordinates from all three subunits (PDB 6V4U or 6LT0) 20,21 , WDR41 (PDB 6WHH), and mouse ARF1 (PDB: 1O3Y) 32 were rigid body fitted separately into the density map using UCSF Chimera 52 . The Leu71 of ARF1 was mutated to its wild-type residue Gln and the GTP was replaced as GDP-BeF 3 − . Atomic coordinates were refined by iteratively performing Phenix realspace refinement and manual building and correction of the refined coordinates in Coot. To avoid overfitting, the map weight was set to 1 and secondary structure restraints were applied during automated real-space refinement. Model quality was assessed using MolProbity 53 and the map-vs-model FSC (Supplementary Table 2 and Supplementary Fig. 6b). A half-map cross-validation test showed no indication of overfitting (Supplementary Fig. 6a). Figures were prepared using UCSF Chimera. The cryo-EM density map has been deposited in the Electron Microscopy Data Bank under accession code EMD-23827 and the coordinates have been deposited in the Protein Data Bank under accession number 7MGE.
High-performance liquid chromatography analysis of nucleotides identity bound to small GTPases. The nucleotides bound to small GTPases were assessed by denaturing 30 μl of protein to 95°C for 5 min followed by 5-min centrifugation at 16,000 × g. The supernatant was loaded onto a HPLC column (Eclipse XDB-C18, Agilent). Nucleotides were eluted with HPLC buffer (10-mM tetra-nbutylammonium bromide, 100-mM potassium phosphate pH 6.5, 7.5% acetonitrile) at flow rate 0.6 ml/min. The identity of the nucleotides was compared to GDP and GTP standards.
HPLC-based GAP assay. HPLC-based GTPase assays were carried out by incubating 30 μl of 30-μM GTPases with or without GAP complex at a 1:100, 1:50, or 1:10 molar ratio as indicated for 15 min at 37°C. Samples were immediately boiled for at least 5 min at 95°C and centrifuged for 5 min at 16,000 g. The supernatant was injected onto an HPLC column as described above. The experiments are carried out at least three times and one representative plot is shown.
Reporting summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article.

Data availability
The data that support this study are available from the corresponding author upon reasonable request. EM density map has been deposited in the EMDB with accession number EMD-23827. Atomic coordinates have been deposited in the PDB with accession number 7MGE.