geometrical restraint among subunits of the
same CR and five rounds of guided multiref-
erence three-dimensional classification. This
practice allowed the selection of 2,477,433 CR
subunits, which yielded a reconstruction of the
core region at 5.77-Å resolution. To fully use the
dataset, these subunits were projected back
to the original CR particles. The defocus values
of all subunit particles within the same CR
were pooled together to calculate a corrected
average defocus value for the center of mass
of each CR particle. All subunits of this CR were
then reextracted using the updated defocus val-
ue deduced from the corrected average defocus
of this CR particle. Through this procedure,
4,528,642 subunit particles from 581,882 CR
particles were extracted, resulting in a recon-
struction at 5.6 Å after autorefinement (fig. S3A).
This dataset was then used for reextraction
with a box size of 400 and pixel size of 1.387 Å
(fig. S3B). Following our published protocol ( 14 ),
data processing beyond this point was per-
formed individually for four datasets defined
by the tilting angles: tilt0, tilt30, tilt45, and
tilt55. This was because of our observation
that reconstruction at relatively high resolu-
tion is easily biased toward low-tilt datasets
due to their relatively high signal-to-noise ratio.
In fact, the high-tilt and low-tilt datasets ex-
hibited quite different motion statistics (fig. S1).
The four datasets were then separately recon-
structed to obtain an estimation of signal-to-noise
ratio for their respective tilting angles. After pol-
ishing and CTF refinements, the four datasets
were pooled for three-dimensional autorefine-
ments and three-dimensional classifications.
A second strategy to enhance the quality of
the reconstruction is to improve the results of
particle polishing. In particle polishing, initial
movement tracks are obtained either through
global frame alignment or polynomial motion
models determined fromlocal motion trajec-
tories ( 43 ).Themagnitudeofthesampledrifts
in the tilted datasets appears much greater
than that in the untilted ones (fig. S1). This
problem, along with increased sample thick-
ness at high-tilt angles, prevents the local
motion model from accommodating sample
deformation. An initial motion model esti-
mation using MotionCor2 resulted in 18,619
failed attempts to explain local sample defor-
mation in 38,583 movie stacks using the
polynomial model. In particular, sample de-
formation was successfully explained for only
6825 movie stacks out of a total of 22,132 for
the tilt45 and tilt55 datasets. Errors introduced
by the polynomial motion model undermine
the accuracy of the initial movement tracks for
particle polishing, thus affecting the final out-
come. To address this issue, we took advantage
of the large size of a single NPC particle and
attempted to estimate only the local motion
around each NPC particle instead of dividing
the micrograph into rectangular patches and
estimating motion for each patch. A major
rationale for this practice is that empty patches,
which occur often, may engender misalignment,
thereby compromising the accuracy of the
polynomial motion model. To improve align-
ment, we used a polynomial motion model
similar to those implemented in MotionCor2
( 41 ), RELION ( 44 ), and WARP ( 45 )toregularize
movement tracks of particles that belong to the
same micrograph. This polynomial model was
iteratively refined until convergence. A final
round of patch-based alignment using only
information around each NPC particle was
performed to account for any residual move-
mentofeachparticle.Finally,anymicrograph
that failed to converge or had residual motion
of >15 pixels in any direction was removed.
Using the above strategy, we were able to
rescue 13,758 micrographs from 18,619 that
failed to provide reliable local motion models.
We separately reconstructed the four tilt datasets
and performed one round of CTF refinement.
The results were directlysubjected to autorefine-
ment, resulting in a reconstruction of the core
region at 4.5-Å resolution (fig. S3B). The com-
bined dataset was then redivided into four
subsets according to tilting angle and sub-
jected to particle polishing and CTF refinement
in RELION. The resulting particles were again
pooled together and subjected to autorefine-
ment and multireference three-dimensional
classifications. The above cycle was repeated
several times. In the last few cycles, magnification
anisotropy refinement and per-micrograph as-
tigmatism refinement were also performed to
reduce distortion due toincorrect magnifica-
tion and astigmatism. Additionally, the box
size of the used particles was enlarged during
particle polishing for the last few cycles using
thewindowandscaleoptionfromrelion_
motion_refine. These measures resulted in a
final average resolution of 3.7 Å from 1,279,270
particles (fig. S3B and table S1).
For the Nup358 region, the particles were
subjected to local refinement in cryoSparc ( 46 )
to reach an average resolution of 4.7 Å (fig.
S3B, and table S1). The EM maps in the core
region display distinct features for secondary
structural elements and some of the bulky
amino acid side chains (figs. S4 and S5). These
features facilitated sequence assignment of
the nucleoporins and identification of protein-
protein interfaces.Expression and purification of Nup358-NTF
The cDNA encoding the N-terminal 1300
amino acids ofX. laevisNup358 (Uniprot:
A0A1L8HGL2) was synthesized with codon
optimization (Qinglan Biotech). The N-terminal
fragment (residues 1 to 1171, referred to as
Nup358-NTF)withanN-terminalFlagtagwas
cloned into the pCAG vector. This construct
was verified by DNA sequencing. HEK293F cells
(Invitrogen) were cultured in SMM 293T-IImedium (SinoBiological) supplemented with
5% CO 2 in a Multitron-Pro shaker at 130 rpm
at 37°C. The cells were transfected at a den-
sity of ~2 × 10^6 cells/ml. SMS 293-SUPI
(SinoBiological) was supplemented into the
culture 24 hours after the transfection. The
cells were then cultured for additional 24 to
36 hours.
The transfected HEK293F cells were har-
vested by centrifugation at 3800g and resus-
pended in a lysis buffer containing 25 mM
Tris-HCl, pH 8.0, 150 mM NaCl, and a cocktail
of protease inhibitors (VWR). The final con-
centrations of the inhibitors were 2 mM
for phenylmethylsulfonyl fluoride (PMSF),
5.2mg/ml for aprotinin, 2.8mg/ml for pep-
statin, and 10mg/ml for leupeptin. The cells
were lysed by ultrasonication (Vibra-Cell,
SONICS). After centrifugation at 30,000g for
1 hour, the supernatant was loaded into an
anti-Flag M2 affinity gel (Sigma-Aldrich) col-
umn and eluted with the lysis buffer sup-
plemented with 200mg/ml Flag peptide. The
eluted fraction was concentrated using a 50-kDa
cut-off Centricon (Millipore) and applied to size-
exclusion chromatography (Superose-6, GE
Healthcare). The peak fractions were ana-
lyzed by SDS-polyacrylamide gel electropho-
resis (fig. S6).Cryo-EM data acquisition for Nup358-NTF
An aliquot of 4ml of freshly purified Nup358-
NTF at a concentration of ~2 mg/ml was
placed on glow-discharged holey carbon grids
(Quantifoil Au 300 mesh, R1.2/1.3). Grids were
blotted for 3.5 s and plunge-frozen in liquid
ethane cooled by liquid nitrogen using Vitrobot
Mark IV (Thermo Fisher) at 8°C under 100%
humidity. The grids were transferred to a Titan
Krios electron microscope (Thermo Fisher)
operating at 300 kV and equipped with a GIF
Quantum energy filter (Gatan). Using the
same setup as that for intactX. laevisNPC data
acquisition, a total of 17,147 movie stacks were
recorded. WARP was used for all subsequent
preprocessing, including CTF estimation and
motion correction ( 45 ).Image processing for Nup358-NTF
Out of a subset of 1176 micrographs, 627,188
particles were first autopicked using Topaz ( 47 ),
with its default model on micrographs pre-
binned by eight times. Particles were extracted
using a box size of 100 and a pixel size of 2.774 Å.
These particles were then subjected to four
rounds of two-dimensional classifications in
cryoSparc ( 46 ), yielding 201,480 particles. The
good particles were then converted to RELION-
compatible star files using the csparc2star.py
program from the pyem suite ( 48 ). These
particles were used to train a deep learning
model in Topaz. Relying on this model, we
picked 2,418,968 particles from 17,147 micro-
graphs using a box size of 100 and a binnedZhuet al., Science 376 , eabl8280 (2022) 10 June 2022 7of10
RESEARCH | STRUCTURE OF THE NUCLEAR PORE