A complete data processing workflow for CryoET and subtomogram averaging

Introduction

• Subnanometer subtomogram averages were demonstrated on previously published datasets, and whole E. coli expressing a double-layer-spanning membrane protein was resolved at 14 Å.The cellular result was obtained with whole E. coli over-expressing the membrane protein.Source: Introduction · p. 2
• The workflow introduces fiducial-free automated tilt-series alignment, tiled direct-Fourier reconstruction, per-particle-per-tilt CTF correction and orientation refinement, and robust initial-model generation.These innovations aim to reduce human intervention and improve the final average’s resolution.Source: Introduction · p. 2
• A graphical interface and cellular-tomography bookkeeping system support analysis of multiple features in one dataset and tracking particles to correlate structures with protein locations.Source: Introduction · p. 2

Automated tilt series alignment and tomogram reconstruction

• Iterative landmark-based alignment works across tomograms with or without fiducials and without human intervention; overlapping-tile Fourier reconstruction significantly reduces edge effects and memory requirements while remaining computationally efficient.Alignment progressively increases sampling as error decreases, refines landmark coordinates and tilt transforms, and excludes a specified fraction of worst-matching landmarks each iteration.Source: Automated tilt series alignment and… · pp. 2–3

Multiple methods for particle localization

• Users can manually pick multiple particle types, use template matching for purified molecules, or convert neural-network annotations into subtomogram coordinates.The tools support both isolated globular particles and continuous structures, with results adjustable through the manual picking interface.Source: Multiple methods for particle… · p. 4

Per-particle-per-tilt CTF correction

• Per-particle CTF correction estimates defocus using information from the full tilt image, enabling correction based on each particle’s position and each tilt.For high tilts, candidate defocus values are constrained to within three standard deviations of the low-tilt mean; corrected subtomograms are reconstructed from fully sampled raw tilt images.Source: Per-particle-per-tilt CTF correction · pp. 4–5

Initial model generation via stochastic gradient descent

• Stochastic gradient descent produces reliable initial models from cell-derived particles, typically within 1 hour on a workstation.Random particle subsets are aligned and averaged iteratively; the reference is low-pass filtered after each update to limit overfitting. The method worked on globular, linear, and double-membrane-spanning structures.Source: Initial model generation via… · pp. 5–6

Subtomogram alignment and averaging

• Purified ribosome resolution improved from 13 Å to 8.5 Å with subtilt refinement; cellular TolC-AcrAB resolution improved from 19 Å initially to 14 Å.The purified result used 3,000 particles from four tomograms; both measurements use the reported resolution values.Source: Subtomogram alignment and averaging · pp. 6–7

Discussion

• The protocol is integrated as a graphical sequence in EMAN2, but advanced methods such as subtilt refinement require following the complete pipeline because of stringent metadata requirements.Some tools, including deep-learning annotation and simple subtomogram alignment and averaging, can still be used on imported data.Source: Discussion · p. 7
• ~10Å resolution is relatively straightforward with 1000 - 2000 particles from a few good tilt series using per-particle CTF correction and subtilt refinement.For phase-plate data, the authors suggest 0.5 - 1 µm underfocus; accurately locating the first CTF zero is sufficient for subnanometer resolution, although jointly determining defocus and phase shift is limited by insufficient information.Source: Discussion · p. 7
• A large soft mask and local-resolution filtering from even and odd sub-maps preserve protein detail between refinement rounds and reduce misalignment from surrounding densities.This addresses the challenge that neighboring strong densities can make masking and filtration strongly affect the final resolution of in situ subtomogram averages.Source: Discussion · pp. 7–8
• 10-15 Å maps can be achieved in days from a few thousand subtomogram particles with per-particle CTF correction and subtilt refinement.Researchers can use manually selected features to automatically annotate similar features, generate de novo initial models without prior protein knowledge, form hypotheses for biochemical validation, and map particle positions and orientations back to tomograms to study cellular organization.Source: Discussion · p. 8

Data availability

• EMD-0529 contains the averaged purified ribosome structure, and EMD-0530 contains the averaged AcrAB-TolC structure from a cellular tomogram.Both subtomogram averages are deposited to the EMDatabank.Source: Data availability · p. 8

Alignment and reconstruction

• Tilt-series alignment progresses from a filtered 512 x 512 coarse alignment to iterative alignment at increasing image sizes, with landmark selection and tomogram reconstruction at each level.Alignments typically proceed through 512, 1024, 2048, and 4096 pixel images; tomograms used for landmark selection are reconstructed from the pre-filtered 512 x 512 tilt series.Source: Methods · p. 15
• The workflow selects tomogram landmarks and jointly refines their 3D locations and tilt parameters, using robust centering and outlier exclusion to limit alignment errors.Refinement proceeds from image translations and global tilt-axis rotation to on- and off-axis tilt angles; the optimizer excludes the 10% highest-error landmarks per tilt, and later excludes the 10% highest-error tilts.Source: Methods · pp. 15–16
• Final tomograms are reconstructed from overlapping tiles with Gaussian-falloff masks to normalize summed voxel weights and reduce edge artifacts.The process takes on the order of 10 minutes per tomogram.Source: Methods · p. 16

Subtomogram averaging

• Initial models are generated by stochastic gradient descent using shuffled particle batches, a low-pass-filtered first-batch average, and iterative reference updates.The default batch size is 12 particles, the learning rate is 0.1, and the initial reference is filtered to 100 Å; the first iteration uses twice as many batches and supports symmetry-axis alignment.Source: Methods · pp. 16–17

Subtilt refinement

• Subtilt refinement searches orientations around each subtomogram orientation, scores projections by CTF-weighted Fourier ring correlation, and weights retained tilt images in the 3D average.Even and odd sets are refined independently; by default, the worst-scoring 50% of particles and scores beyond 2 standard deviations are removed, and the average is filtered by gold-standard FSC.Source: Methods · p. 17

Processing of example data sets

• Example processing applied the workflow to purified ribosome particles and cellular AcrAB-TolC, with different initialization strategies for the ribosome and symmetric pump.The ribosome workflow used 3,239 particles, a 50 Å initial-model target, four subtomogram-refinement rounds, and three subtilt-refinement rounds. For AcrAB-TolC, 25 particles seeded C1 then C3 model generation before averaging 1,321 particles from nine tomograms with C3 symmetry.Source: Methods · pp. 17–18

• Overlapping tiles reduce edge effects in tomogram reconstruction, with each tile padded during reconstruction and clipped to its target size.The resulting reconstruction is assembled by tile.Source: Figure S1 · p. 19
• Per-particle defoci are determined from strip-wise coherent power spectra, tilt-angle geometry, and particle positions, then fitted CTF parameters phase flip subtilt images.CTF is measured in each tilt image by tiling it and analyzing strips parallel to the tilt axis.Source: Figure S2 · p. 20