REVIEW 2 major objections 4 minor 4 references
Resolving structural dynamics in situ through cryogenic electron tomography
T0 review · 2 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read This review argues that cryo-ET structural dynamics analysis is moving from 3D subtomogram volumes to 2D particle-image stacks, and that the field's next bottleneck is benchmarking datasets with per-particle ground truth.
desk verdict A solid, useful survey of cryo-ET heterogeneity workflows whose central 'shift to 2D' claim is stronger than its own evidence; the benchmarking advocacy is the strongest part. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The organizing device is a two-axis taxonomy: data representation (3D subtomogram volumes versus stacks of 2D tilt-series images) and heterogeneity model (discrete classes versus a continuous latent space). Along the continuous axis, the review highlights autoencoder architectures as the mechanism that carries heterogeneous reconstruction: an encoder compresses each particle into a low-dimensional latent variable, a decoder reconstructs the input, and training is self-supervised by minimizing reconstruction error. This lets each particle occupy its own point in latent space, so rare states and continuous transitions are not forced into a fixed number of classes.
What would settle it
A head-to-head evaluation in which 3D subtomogram-based methods consistently recover conformations at higher resolution or classify rare states more accurately than 2D particle-stack methods on the same benchmark data would undercut the claimed shift to 2D workflows.
Extended reading notes
Core claim
The field has recently shifted away from 3D subtomograms and towards using 2D particle-stacks, an approach that reduces storage and computational costs while enabling higher-resolution reconstructions. By operating directly on tilt-series images, 2D workflows allow fine-grained, per-particle refinement of tilt geometry, motion, and CTF parameters during reconstruction, capabilities that are difficult to implement with interpolated 3D subtomograms. The review's claim is that this shift, combined with continuous classification methods that assign each particle a point in a learned latent space, is what now allows researchers to resolve discrete structural states and continuous conformational changes in situ. The paper further argues that the field's progress is limited less by algorithms than by the lack of benchmarking datasets that provide per-particle ground truth, and it points to a programmable in vitro dataset as a template for such benchmarks.
Load-bearing premise
The paper's central proposal depends on the feasibility of benchmarking datasets in situ, but for particles inside cells there is currently no independent method that establishes per-particle conformational or compositional ground truth.
Editorial extensions
If this is right
- Existing 3D subtomogram workflows will increasingly be ported to or replaced by 2D particle-stack pipelines.
- Continuous classification in latent space will become a standard first pass for exploratory analysis of in situ heterogeneity.
- Benchmarking datasets with per-particle ground truth will become necessary for choosing among methods, similar to benchmarks in single-particle cryo-EM.
- The TRiC chaperonin example shows that in situ heterogeneous reconstruction can already expose cofactor occupancy and open/closed states, and such analyses should become more common.
Reading between the lines
- The review's strongest implicit wager is that 2D-based methods will outperform 3D-based methods on most future benchmarks, but this is not yet demonstrated by experiments.
- True per-particle ground truth inside cells may be unobtainable by independent means, so the most decisive benchmarks may be hybrid ones that mix known compositions into a cellular context.
- Latent-space methods could be extended to predict not just density but functional labels, such as ligand occupancy or interaction partners, if benchmarks provide such annotations.
- A testable prediction: methods that use all tilt images, weighted by signal-to-noise ratio and dose, will show larger gains on thick or crowded cellular samples than on purified in vitro particles.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This review surveys computational methods for analyzing structural heterogeneity in cryo-electron tomography (cryo-ET), organizing the field along two axes: workflows that operate on 3D subtomograms versus 2D particle-image stacks, and methods that perform discrete versus continuous classification. It presents a meta-analysis of EMDB deposits (Figure 1A), a comparative table of twelve tools (Table 1), and a discussion of recent biological applications. The paper concludes by advocating for the development of benchmarking datasets, both in vitro and in situ, to evaluate and motivate future methodological innovations.
Significance. If its characterization of the field is accurate, the review provides a timely and useful synthesis of a rapidly moving area, and the comparative table is a convenient resource for practitioners. The emphasis on benchmarking is constructive and timely. The paper is clearly written and the meta-analysis of EMDB growth is a nice contribution. However, the central narrative claim that the field has 'recently shifted' from 3D subtomograms to 2D particle stacks is not substantiated by the evidence presented, and the paper's own Table 1 actually shows a nearly balanced or even 3D-leaning set of recent tools. This weakens the authority of the review's framing and the motivation for its benchmarking recommendations, though the claim is potentially true and could be fixed with a systematic census or a more measured framing.
major comments (2)
- [Structure determination using 2D particle-image stacks or 3D subtomogram volumes] The claim that 'The field has recently shifted away from 3D subtomograms and towards using 2D particle-stacks' is a factual assertion about community practice that is not supported by the paper's evidence. Figure 1A, the EMDB meta-analysis, contains no field indicating workflow type and therefore cannot distinguish between studies that used 2D particle stacks and those that used 3D subtomograms. Table 1, the paper's only other systematic survey, actually lists seven tools operating on 3D volumes (RELION 3.1-4.0, STOPGAP, OPUS-TOMO, Dynamo, PEET, MDTOMO, TomoFlow) against five operating on 2D stacks (RELION 5.0, NextPYP, tomoDRGN, cryoDRGN-ET, emClarity), and several recent entries (OPUS-TOMO, STOPGAP in 2024; MDTOMO in 2023) fall on the 3D side. The claim may be true, but it requires a systematic census of published cryo-ET workflows (e.g., by year and by processing strategy) before it can be accepted as stated. Because the review's organizational framing and its benchmarking advocacy are motivated by this shift narrative, the unsupported version of the claim is a load-bearing weakness. The manuscript should either supply such a census or reframe the comparison as one between two coexisting strategies.
- [What lies ahead: the need for benchmarking datasets] The statement that 'Acquiring tilt-series data of this type should be feasible, and future efforts to generate analogous datasets in cellular cryo-ET could serve as definitive benchmarks' overstates the feasibility of ground-truth benchmarking in situ. Per-particle conformational or compositional ground truth inside cells cannot currently be established by an independent experimental method; in vitro benchmarks can control composition but do not capture the cellular environment that in situ cryo-ET is designed to study. The manuscript should explicitly acknowledge this limitation and discuss what independent validation (e.g., cross-correlation with orthogonal imaging or synthetic ground truth) would be needed for such datasets to serve as 'definitive' benchmarks.
minor comments (4)
- [References] The reference for LeCun et al. 1998 is malformed; it begins 'Y, Bottou L' instead of 'LeCun Y, Bottou L'.
- [General] The running header on the first pages ('Resolving structural heterogeneity in cells with cryo-ET') differs from the paper's title and appears to be a leftover from an earlier version.
- [Figure 2] The caption says the schematic is 'Adapted from (Powell and Davis 2024)' but does not indicate what modifications were made; please specify.
- [Table 1] The binary categorization into '2D' and '3D' obscures that some tools (e.g., RELION 5.0) support both modes or that '2D' tools still reconstruct 3D volumes; a brief note in the caption or text acknowledging this spectrum would improve accuracy.
Circularity Check
No significant circularity: the review's claims are external or programmatic, not derived from the authors' own tools.
full rationale
This is a narrative review, not a derivation. Its only quantitative evidence, the EMDB meta-analysis in Figure 1A, is an external dataset retrieved from the Electron Microscopy Data Bank and annotated under the method 'subtomogram averaging.' The comparative claim that the field has 'recently shifted away from 3D subtomograms and towards using 2D particle-stacks' is a community-practice assertion supported by cited software descriptions (Warp, M, RELION-5, NextPYP) rather than by the authors' equations; it may be contestable as a factual matter, but it is not circular. The self-citations to tomoDRGN, cryoDRGN, SIREn, and the Kinman benchmarks appear in descriptive passages about autoencoder architectures and in the benchmarking proposal; they do not function as premises that force a conclusion. The final benchmarking recommendation is explicitly framed as an advocacy statement ('we advocate for the development of benchmarking datasets'), not as a result derived from fitted parameters or from the authors' prior work. No equation, fitted parameter, or imported uniqueness theorem is invoked, so no circular step can be exhibited.
Assumptions & free parameters
assumptions (3)
- domain assumption EMDB entries annotated with the method 'subtomogram averaging' provide a complete and unbiased census of relevant structures.
- domain assumption The primary literature accurately describes each tool's capabilities and limitations.
- domain assumption The categorical axes (2D vs 3D, discrete vs continuous) are meaningful and mutually exclusive for classifying methods.
Cite this review
Pith. "Pith review of Resolving structural dynamics in situ through cryogenic electron tomography." pith.science (2026). https://pith.science/paper/2UHH46RG
@misc{pith2026250622719,
author = {Pith},
title = {Pith review of: Resolving structural dynamics in situ through cryogenic electron tomography},
year = {2026},
howpublished = {\url{https://pith.science/paper/2UHH46RG}},
note = {Machine review of arXiv:2506.22719}
}
read the original abstract
Cryo-electron tomography (cryo-ET) has emerged as a powerful tool for studying the structural heterogeneity of proteins and their complexes, offering insights into macromolecular dynamics directly within cells. Driven by recent computational advances, including powerful machine learning frameworks, researchers can now resolve both discrete structural states and continuous conformational changes from 3D subtomograms and stacks of 2D particle-images acquired across tilt-series. In this review, we survey recent innovations in particle classification and heterogeneous 3D reconstruction methods, focusing specifically on the relative merits of workflows that operate on reconstructed 3D subtomogram volumes compared to those using extracted 2D particle-images. We additionally highlight how these methods have provided specific biological insights into the organization, dynamics, and structural variability of cellular components. Finally, we advocate for the development of benchmarking datasets collected in vitro and in situ to enable a more objective comparison of existent and emerging methods for particle classification and heterogeneous 3D reconstruction.
Figures
Reference graph
Works this paper leans on
-
[4]
Reconstructing continuous distributions of 3D protein structure from cryo-EM images
Cryo-electron tomography reveals the structural diversity of cardiac proteins in their cellular context. bioRxiv 2023.10.26.564098. Xing H, Rosenkranz RRE, Rodriguez-Aliaga P , Lee T-T, Majtner T, Böhm S, Turoňová B, Frydman J, Beck M. 2025. In situ analysis reveals the TRiC duty cycle and PDCD5 as an open-state cofactor. Nature 637: 983–990. Xing H, Tani...
work page Pith review arXiv 2023
-
[2008]
CryoBench: Diverse and challenging datasets for the heterogeneity problem in cryo-EM
Classification and 3D averaging with missing wedge correction in biological electron tomography. Journal of Structural Biology 162: 436–450. Basiashvili T, Kim S, Chen S, Hutchings J, Reck-Peterson SL, Villa E. 2023. Molecular architecture of LRRK2 in association with cellular organelles. Biophysical Journal 122: 544a. Baumeister W. 2022. Cryo-electron to...
work page Pith review arXiv 2023
-
[2023]
Journal of Structural Biology 215: 107911
Multishot tomography for high-resolution in situ subtomogram averaging. Journal of Structural Biology 215: 107911. Khusainov I, Romanov N, Goemans C, Turoňová B, Zimmerli CE, Welsch S, Langer JD, Typas A, Beck M. 2024. Bactericidal effect of tetracycline in E. coli strain ED1a may be associated with ribosome dysfunction. Nature Communications 15: 4783. Ki...
arXiv 2024
-
[2025]
In-cell structure and snapshots of copia retrotransposons in intact tissue by cryo-electron tomography. Cell 188(8):2094-2100. Y, Bottou L, Bengio Y, Haffner P . 1998. Gradient-based learning applied to document recognition. Proceedings of the IEEE 86: 2278–2324. Li S, Fernandez J-J, Marshall WF, Agard DA. 2019. Electron cryo-tomography provides insight i...
work page 1998
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.