Cell culture and Nup96 depletion
NUP96::Neon-AID DLD-1 cells23 were grown in standard tissue culture conditions (37 °C with 5% CO2) using DMEM medium supplemented with 4.5 g l−1 glucose and 4.5 g l−1 sodium pyruvate (Corning), 10% FBS (Gemini Bio) and 2 mM GlutaMAX (Gibco). Next, cells were cultured for 18–24 h on freshly glow-discharged carbon-coated EM gold grids (200 mesh, R1.2/1.3 or R2/1; Quantifoil Micro Tools) before plunge-freezing. For Nup96 depletion, a 250 mM auxin (3-indoleacetic acid, Sigma-Aldrich) solution was added to the cells to a final concentration of 1 mM, followed by incubation at 37 °C for 4–8 h before plunge-freezing. HeLa-derived (TOR1A/TOR1B/TOR3A-knockout) cells43 were cultured in the same conditions but never subjected to auxin treatment.
Cryo-FIB milling of lamellae from DLD-1 cells
To prepare lamellae, we used both an Aquilos FIB-SEM system (Thermo Fisher Scientific) and a Crossbeam 540 (Zeiss) with cryo stage (Leica) using similar methods as previously described44. On the Aquilos system, grids were loaded into the sample chamber and sputtered with an initial platinum coat (15 s) followed by a second organometallic platinum protective layer using the gas injection system (GIS, 15 s). Samples were tilted to an angle of 15°, and FIB ablation of cellular material was performed in a stepwise manner by focusing the gallium beam at 30 kV on parallel rectangular patterns with over- and under-tilting to the desired thickness as follows: (1) 15°, 500 pA, gap = 3 µm; (2) 16°, 300 pA, gap = 2.2 µm; (3) 14°, 300 pA, gap = 1.5 µm; (4) 15.5°, 100 pA, gap = 0.8 µm; (5) 14.5°, 100 pA, gap = 0.5 µm; (6) 15°, 50 pA, gap = 300 nm; and (7) 15°, 30 pA, gap = 150 nm. The initial width of the lamellae was set to 13 µm but, at each reduced current, we reduced the width by 0.25 µm, for lamellae with a final width of 11–12 µm. Expansion/relief joints to reduce tension at the lamella were milled 3 µm away from the sides of the lamella at step 1 with a 1-µm-wide pattern. A similar procedure was used on the Crossbeam 540 FIB-SEM. Grids were first electron-beam-coated with 2 nm of platinum in an ACE 900 system (Leica) before being loaded into the Crossbeam sample chamber with a VCT 500 (Leica) vacuum transfer system for organometallic platinum coating. FIB ablation was performed using only four stepwise currents and without over- and under-tilting as follows: (1) 700 pA, gap = 3 µm; (2) 300 pA, gap = 1 µm; (3) 100 pA, gap = 400 nm; and (4) 50 pA, gap = 150 nm. In total, we prepared 37 lamellae from non-auxin-treated NUP96::Neon-AID DLD-1 cells, and 115 lamellae from auxin-treated cells to obtain our data.
Tilt series acquisition and processing
Cryo-ET datasets were acquired on a Titan Krios G3i operating at 300 keV equipped with a BioQuantum post-column energy filter and a Gatan K3 direct electron detector. Tilt series were acquired at a magnification of ×26,000, resulting in a pixel size of 3.4 Å at the specimen level using Tomo v.5.3.0 (Thermo Fisher Scientific). We used a tilt range of −52° to +68°, starting at +8° with a bidirectional scheme45 resulting in a total dose of about 145 electrons per angstrom squared. Tilt series of HeLa cells (TOR1A/TOR1B/TOR3A-knockout)43 were acquired with an accumulative dose of about 120 electrons per angstrom squared. Defocus values varied between −5.0 µm to −2.5 µm through the entire data acquisition process. Representative tomograms are included (Supplementary Videos 1 and 2).
Tilt series were aligned with 4× binned projections using patch tracking in the IMOD software package46. When available, contaminations generated by the FIB process were used as fiducial markers. The contrast transfer function (CTF) was determined and corrected as described previously26, and is summarized here. The mean defocus was estimated by strip-based periodogram averaging for each tilt series. Using the mean defocus, the tilt angle and axis orientation, the defocus gradient for each projection was determined. CTF correction using the defocus gradient was then performed by phase-flipping each projection image. CTF-corrected stacks were next dose-filtered using the IMOD mtffilter function46. NPC coordinates were picked manually using IMOD. During NPC particle picking, the relative orientation of the NPC to the NE was determined to prealign the NPCs. Next, subtomograms of NPCs were reconstructed using IMOD. Detailed imaging parameters are summarized in Extended Data Table 1.
For the wild-type DLD-1 structure, 194 prealigned NPCs were further aligned using iterative missing-wedge-weighted subtomogram alignment and averaging with the TOM toolbox (tom_corr3d)47. Half-set averages were merged after each iteration and used as the template for the next iteration. First, full NPCs were aligned using 8× binned subvolumes and applying eightfold-fold symmetry. This step was repeated using 4× binned subvolumes. Subsequently, these alignments were used to extract the eight asymmetric units (1,552 protomers) for each NPC according to the eight-fold symmetry as previously described48. Twice-binned full protomers were aligned and further refined by applying masks for the individual rings (CR, IR and NR). Next, subprotomers were extracted based on the previous alignment step. At this stage, single protomers were manually inspected and missing protomers from incomplete NPCs, misaligned protomers and protomers with a low signal-to-noise ratio were excluded. After this step, 1,252 protomers remained for the subsequent steps. Newly extracted subprotomers were further aligned. Resolution was measured using the 0.5 criterion and soft masks to exclude an artificial contribution to the measured resolution using the Electron Microscopy Data Bank (EMDB) validation server (https://www.ebi.ac.uk/pdbe/emdb/validation/fsc/). Subprotomer volumes were B-factor sharpened using Relion49 with a B-factor of −2,000 Å2. The final full NPC model was generated by fitting the subprotomer volumes into the full NPC average using UCSF Chimera50.
Averaging of auxin-treated (Nup96-depleted) data was performed similarly. First, 163 full NPCs were aligned using 8× binned subprotomers and applying eight-fold symmetry. Next, NPCs were classified by manually inspecting the subvolumes into three classes containing CR, IR and NR (27 NPCs), IR and NR (53 NPCs), or IR only (83 NPCs). These three classes were further aligned using 8× binned subvolumes at full NPC level using eight-fold symmetry.
For the HeLa (TOR1A/TOR1B/TOR3A-knockout) average, 27 NPCs were aligned similar to the wild-type dataset but without dose-filtering of the stacks. After protomer extraction, 36 protomers were excluded and protomers were aligned. The final map (180 protomers) was generated by fitting the protomer average back into the full NPC map using Chimera.
To create models for the Y complexes in the CR and NR, we performed unbiased global fitting using structural models derived from previously reported human NPC structures9,15. For these complexes, we began with the previously reported model and performed a global fitting analysis as implemented in UCSF Chimera50. The fitting was performed independently for the CR and NR using a three-protomer segmented model as our single protomers do not necessarily contain a complete protomer of the eight-fold symmetric NPC (that is, a single protomer may contain half of the Y complex from adjacent protomers). For the IR, a single protomer was used as it contained the entire complex. All fitting runs were performed using Chimera and 1,000,000 random initial placements and local cross-correlation (Chimera’s correlation about the mean (CAM)), or a combination of local cross-correlation and overlap (CAM + OVR) was used when local cross-correlation did not provide statistically significant fits (as others have observed)20. We computed fit scores using both metrics for each model but, for the IR, the local cross-correlation (CAM) provided statistical significance to our model, whereas, for the Y complex, the combination metric (CAM + OVR) provided statistical significance. We noted the scores and adjusted P values (described below) on the figure for each scoring metric. For each fitting run, the statistical significance was assessed as a P value that was calculated from normalized fitting scores. To calculate the P values, we transformed the CAM or combined CAM + OVR scores into Z scores, derived a two-sided P value for each fit and then corrected the P values for multiple testing using the conservative Benjamini–Hochberg procedure. This workflow was performed using a Python script running SciPy.Stats (for P value and Z-score analysis)51, the StatsModels module (for Benjamini–Hochberg analysis)52 and Matplotlib (for plots)53.
For the IR complex, we identified a significant fit in the expected position as described in the most recent human structure (Protein Data Bank (PDB): 5IJN)9, although it was obvious that some of the Nups in this model could be shifted to better fit our density map. We therefore followed a previously described approach20 and optimized the fits by local refitting of individual subunits or domains. For this refinement, we used the option in Chimera to use a map simulated from atoms at a resolution of 25 Å and optimized the fitting for correlation. Using this approach, we fit subcomplexes made of the chains F, G, H, X, Y and Z (two copies of Nup54, Nup58 and Nup62), L, M, N, R, S and T (two additional copies of Nup54, Nup58 and Nup62), each copy of Nup205 (chains D, J, P and V), and the four copies of Nup93 (chains C, I, O and U). To fit the Nup155 chains, we subtracted densities with a radius of 10 Å of the fitted models above from the determined NPC map. In this difference map, we fit the individual Nup155 copies into the map using the procedure described above (chains E, Q, K, W). The Nup155 chains A and B of 5IJN could not be fit into our map reliably. We next ran a new unbiased global search using this model as the input for completeness.
For the Y complex, we identified a significant fit in the same position as described in the most recent human NPC structure CR and NR (PDB: 5A9Q)15 using a double Y complex as the reference molecule. As the 5A9Q model contains gaps in the heterotetrameric core element of the Y complex called the ‘hub’ (Nup160, Nup85, Nup96 and Sec13), we decided to replace the hub of the 5A9Q model with the hub of a new composite model that we created using published crystal structures and threading models. We first generated an S. cerevisiae composite model combining a previous structure of the Y complex hub (Nup120–Nup145N–Sec13–Nup85)14 (PDB: 4XMM) and the Nup84–Nup133 subcomplex24 (PDB: 6X02). This S. cerevisiae Y-complex composite was then used to template the human homologue, using published structures and threading models. The hub of this updated model was then used to rigid-body dock into the CR/NR of our map at the same position that the 5A9Q hub occupied, and the Nup107–Nup133 subcomplex from 5A9Q was kept as is, creating a new complete Y-complex assembly. Fitting was further refined by cutting Nup160 between residue 933 and 934 (ref. 11) and fitting the N-terminal part of Nup160 (Nup160-N) together with Nup37 as a rigid body into our map. The C-terminal part of Nup160 was fit as part of the hub. The CR model was then also placed for the inner Y ring and not further modified. Fitting for the NR Y complex was performed similarly. For the outer Y ring, only the hub and the Nup107–Nup133 subcomplex were separated and fit as individual rigid bodies. For the IR of the NR Y complex, the model without Nup160-N and Nup37 was fit into the map. For fitting of the Nup160-N–Nup37 subcomplex, we subtracted the densities of previously fit models with a radius of 20 Å from the map and fit the subcomplex into the remaining density. After creating the final models for Y complexes in both the CR and NR, we performed another round of unbiased global fitting and identified a similar location at which the Y complex of 5A9Q docked. A final local optimization run maximized density overlap. We strongly urge that the final fits are not interpreted at atomic resolution. Instead, our fitting simply aids in assignment of our density toward understanding how each of the subcomplexes is positioned.
The IR model described above was placed manually into the centre of the IR protomers of the subtomogram maps generated from auxin-treated (Nup96-depleted) cells.
Visualizations were performed using UCSF Chimera50. Representative tomograms of DLD-1 cells were reconstructed in 8× binned, 12× SIRT-like filtered tomograms in IMOD46. Snapshots of mouse embryonic fibroblast (MEF) NPCs were recorded after manually aligning the NPCs in IMOD Slicer. The orthoslice views of the full-NPC average, wild-type map and auxin maps were taken using tom_volxyz. Scripts were implemented in MATLAB (MathWorks) and using the tom_toolbox54.
Local resolution analysis was performed as previously reported26, but also summarized here. To calculate the local resolution of the subprotomers, the full subprotomer volume (100 × 100 × 100 voxels) was divided into smaller subvolumes (box size, 40 × 40 × 40) along a regular spacing of 4 × 4 × 4. Resolution was measured between the subvolumes of the two half sets and using the 0.5 threshold criterion. Subvolumes were masked with a spherical mask prior to FSC calculation. Data points between measured values were interpolated and visualized using Chimera’s Surface color function.
NPC diameter analysis
For diameter measurements, single NPCs were measured using orthoslices at the level of the pore membrane. For the DLD1 and HeLa datasets, aligned NPCs that would proceed to full-NPC averaging in this study were used. For the MEF datasets29, NPCs were manually aligned in IMOD slicer and measured. Distance was measured between the NE at the narrowest point of each NPC manually. When possible, measurement was performed in two orthogonal directions and the average was calculated. Otherwise, only a single measurement was performed per NPC. When measurement was not possible because of strong misalignment or a poor signal-to-noise ratio, no diameter was measured. Scatter plots for NPC diameter analysis were created using Prism 9 (GraphPad).
Statistics and reproducibility
Representative micrographs are provided in Fig. 1a and Extended Data Fig. 6a, f. The micrograph in Fig. 1a highlights the three-ring NPC architecture directly visualized in cryo-ET and was chosen from the dataset of 54 wild-type DLD-1 cell tomograms, all of which reproducibly show a three-ring architecture. The micrograph in Extended Data Fig. 6a was chosen from the auxin-depleted DLD-1 cell dataset (71 tomograms), and this image was specifically chosen because it highlights single-ring NPCs that we describe in this Article. Finally, the micrograph in Extended Data Fig. 6f was specifically chosen to provide the reader with an anecdotal observation that we identified only twice in the dataset (2 out of 73 tomograms) and these tomograms were excluded from subsequent processing (71 tomograms were used for subtomogram averaging).
Further information on research design is available in the Nature Research Reporting Summary linked to this paper.