Pith. sign in

REVIEW 4 major objections 4 minor 49 references

Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery

T0 review · 4 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A hierarchical Gaussian mixture with discovered parts tops heterogeneous cryo-EM reconstruction.

desk verdict CryoSPIRE is a genuine GMM-based advance on CryoBench with a novel part-discovery idea, but the part discovery is only qualitatively validated and the abstract's state-of-the-art claim overstates the Ribosembly result. read the letter →

arxiv 2506.09063 v1 pith:DLPWSWLU submitted 2025-06-06 q-bio.QM cs.CVcs.LGeess.IV

classification q-bio.QMcs.CVcs.LGeess.IV PACS 87.64.Ee
keywords cryo-EMheterogeneityGaussianmixturemodelpartdiscoveryconformationalvariabilitycompositionalhierarchicalreconstructionCryoBenchPer-ConformationFSC
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces CryoSPIRE, a two-stage reconstruction framework for cryo-electron microscopy that models a biomolecule as a hierarchy of Gaussian components anchored to parts. The first stage learns a coarse Gaussian mixture whose per-component feature vectors, after clustering, divide the molecule into parts that move or appear together. The second stage uses those parts as anchors to fit a high-resolution mixture with per-image latent codes controlling rigid-body anchor motion, local Gaussian deformation, and on-off amplitude modulation. On the CryoBench synthetic benchmark, CryoSPIRE reports the highest mean Per-Conformation FSC AUC on the IgG-1D and IgG-RL datasets, matches the top method on Ribosembly, and is the first GMM-based method benchmarked there. If correct, this shows that discovering a part-based inductive bias from data is enough to handle both conformational flexibility and compositional variation in one model.

What carries the argument

The central object is the hierarchical anchor-Gaussian mixture: a set of anchors, each carrying a center and a feature vector, with Gaussians attached to anchors through center and feature offsets; per-image latent codes condition three MLPs that generate anchor rigid-body motions, per-Gaussian residual shifts, and sigmoid amplitude modulation for occupancy. The load-bearing mechanism is part discovery: k-means++ clustering on the learned feature space, followed in some datasets by spatial clustering, assigns each Gaussian to a part. In one phrase, the identity is learnable feature-conditioned piecewise-smooth deformation with occupancy modulation.

What would settle it

Retrain the part-discovery stage from several random initializations and check whether the same parts emerge; then train the second stage with part assignments randomly permuted while keeping the same anchor count and loss. If the CryoBench AUC-FSC does not drop when the discovered parts are replaced by random groupings, the part-based inductive bias is not what the performance gain rests on.

Watch

Extended reading notes

Core claim

The core claim is that a coarse Gaussian mixture with learnable per-component features has feature structure that mirrors structural parts: components that co-move or co-appear converge to nearby feature vectors without any supervision on features. Clustering those features yields a part segmentation, and a second-stage model that ties Gaussian motion to one anchor per part learns locally rigid, composable deformations. On CryoBench, this outperforms widely used latent-variable and deformation baselines, sometimes by a wide margin, and on experimental ribosome and spliceosome data it recovers parts that match previously reported structural domains.

Load-bearing premise

The load-bearing premise is that after feature clustering, each anchor groups Gaussians that truly move and appear together; if the clusters are not coherent kinematic units, the hierarchical motion model is built on a faulty inductive bias.

Editorial extensions

If this is right

  • A single framework now covers both continuous conformational changes and discrete compositional changes, so one does not need separate deformation and classification pipelines.
  • Part-based anchors give a route to interpreting latent trajectories as large-scale rigid motions of domains, which is closer to how structural biologists describe complexes.
  • Because Gaussian components are sparse and atomic, the output can feed downstream molecular fitting or physics-based priors, extending earlier GMM work.
  • The ablation implies that replacing the learned feature space with positional encodings, as in DynaMight, costs accuracy on CryoBench, so feature learning rather than coordinate smoothness is what the gain hinges on.
  • On Ribosembly, where linear subspace models are favored by design, CryoSPIRE matches but does not clearly beat RECOVAR, so the method's edge is largest when motion dominates.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the feature-clustering step could be turned into a quantitative part-validity measure, for example by comparing discovered parts against independently known structural domains on a curated set of complexes, rather than relying only on visual inspection.
  • Beyond the paper: one could track whether cluster assignments stay stable across different random initializations of the part-discovery stage; stable parts would strengthen the claim that features encode structural regularities.
  • Beyond the paper: because anchors persist across latent states, the occupancy activations of parts could be used to estimate per-state populations directly, giving a compositional state count without supervised classification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper introduces CryoSPIRE, a hierarchical Gaussian mixture model for heterogeneous cryo-EM reconstruction. The method first trains a coarse-grained GMM with learnable per-Gaussian features, then clusters those features (with k-means++) to discover a part-based segmentation. Each part becomes an anchor in a hierarchical model where the latent code per image drives rigid anchor motions, per-Gaussian displacements, and amplitude modulations for compositional variability. The model is evaluated on the CryoBench synthetic benchmark (IgG-1D, IgG-RL, Ribosembly) and on two experimental datasets (EMPIAR-10076, EMPIAR-10180). The authors report state-of-the-art AUC-FSC on CryoBench compared with RECOVAR, 3DVA, 3DFlex, CryoDRGN, and DRGN-AI, and they provide ablation studies supporting the hierarchical motion model, the learnable Gaussian features, and the importance of the discovered parts.

Significance. If the results hold, CryoSPIRE provides a principled and interpretable way to inject part-based inductive bias into heterogeneous cryo-EM reconstruction. The use of CryoBench, a standardized benchmark with ground-truth states, is a clear strength, as is the inclusion of ablations that isolate key design choices. The paper is also, to my knowledge, the first GMM-based method to be benchmarked on CryoBench. However, the central novelty — automatic part discovery — is only validated qualitatively, and the number of parts is selected manually. Moreover, the state-of-the-art claim is not uniform across all CryoBench datasets: on Ribosembly the gap to RECOVAR is not statistically significant. These issues do not invalidate the approach, but they need to be addressed before the paper's strongest claims are fully supported.

major comments (4)
  1. [Sec. 3.2 and Supplement C] The claim that learned Gaussian features 'naturally group Gaussians into 3D parts' is supported only by UMAP visualizations and qualitative figures; no quantitative comparison is made between the inferred parts and ground-truth structural parts on CryoBench. Supplement C explicitly states that the number of anchors is chosen manually, e.g., using UMAP visualization. This makes the part-discovery step manual rather than fully automatic, and it means that the reported benchmark gains could depend on the selected anchor count. I recommend adding a quantitative evaluation of part discovery (for example, measuring overlap or consistency between inferred clusters and the known structural states on CryoBench) and either automating the selection of the number of clusters or demonstrating robustness across a range of plausible anchor counts.
  2. [Sec. 5, Table 1] The Abstract and Introduction state that CryoSPIRE establishes a new state-of-the-art on CryoBench, but on Ribosembly CryoSPIRE (0.427 ± 0.014) is statistically indistinguishable from RECOVAR (0.430 ± 0.016), as the text itself acknowledges. Since Ribosembly is one of the three benchmark datasets, the state-of-the-art claim should be qualified: either restrict the claim to the IgG datasets or report significance tests across independent training runs to support the comparison on all datasets.
  3. [Sec. 5.1, Table 2] The ablation study shows that over-segmenting to 64 anchors reduces AUC on all datasets, and that removing hierarchical motion also reduces performance. Together with the manual anchor selection, this sensitivity indicates that the benchmark numbers may be specific to the chosen anchor configuration rather than to the proposed part-discovery mechanism per se. The authors should report performance across a sweep of anchor counts or provide a data-driven criterion for selecting the number of anchors, so that the results can be attributed to the method rather than to manual tuning.
  4. [Sec. 4 and Table 1] The reported standard deviations in Table 1 are computed across structural states, not across independent training runs. Because the optimization involves stochastic initialization and sampling, the claim that CryoSPIRE 'significantly outperforms' RECOVAR on IgG datasets is not supported without multiple seeds and an appropriate paired significance test. Please report mean and standard deviation over at least three independent training runs, or otherwise justify the current statistics.
minor comments (4)
  1. [Abstract and Supplement A] The paper refers to a 'Project Webpage' but does not provide a URL. Since the qualitative comparisons and videos are said to be on the webpage, the URL should be included for reproducibility.
  2. [Fig. 7 and Sec. 5] The text refers to 'Fig. 7E', but Figure 7 contains panels (A)-(C); this should be corrected to 'Fig. 7C'.
  3. [Throughout] The abbreviation 'c.f.' should be 'cf.' (for example, in Section 3.2). There are also a few grammatical issues, such as 'the structure of the model is grounded in an initial process that infers a part-based segmentation' being clearer as 'the model structure is grounded in an initial process that infers a part-based segmentation.'
  4. [Sec. 4] The definition of AUC for FSC curves is implicit. Please state explicitly that the area under the FSC curve is computed up to the Nyquist frequency (or the maximum frequency sampled by the grid), so that the metric is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CryoBench SOTA claim is evaluated against external ground-truth benchmark data, and the part-discovery step is an unsupervised initialization whose manual anchor choice is model selection, not a fitted prediction.

full rationale

The paper's central quantitative claim is benchmarked on CryoBench, an external benchmark with ground-truth structural states; the AUC-FSC values in Table 1 measure agreement between model-generated density maps and ground-truth states, and no benchmark label is used in fitting the model. The part-discovery stage is an unsupervised k-means clustering of learned Gaussian feature vectors, and although the number of anchors is chosen manually (Supplement C) and the resulting part segmentation is not quantitatively validated against ground-truth parts, this is a validation/model-selection gap rather than a circular reduction: the clustering is not defined in terms of the benchmark metric, the anchor hierarchy is only one component of the full model, and the ablations show that performance is sensitive to whether hierarchical motion or the chosen number of anchors is used. The authors' prior work (3DVA, 3DFlex, CryoSPARC) appears as baselines and preprocessing tools, not as the justification for the central claim. Limitation statements in Sec. 6 acknowledge open problems in validating experimental latent spaces and in requiring initial poses/structure, but these are honest caveats, not circular derivations. No equation reduces to its own input, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 8 free parameters · 6 assumptions · 1 invented entities

The model's correctness rests on the Gaussian image formation model, external pose estimation, the unsupervised feature-clustering assumption, and several hyperparameters chosen by hand. No new physical entities are introduced; 'anchors' are computational constructs.

free parameters (8)
  • Number of anchors per dataset = 5 (IgG-1D), 6 (IgG-RL), 8 (Ribosembly), 4 (EMPIAR-10076), 4 (EMPIAR-10180)
    Manually chosen by authors via UMAP visualizations of the feature space (Supplement C). This hand selection directly influences the part decomposition and downstream reconstruction.
  • Number of Gaussian components G = 2048 (part discovery); 8192 or 16384 (part-aware)
    Set per dataset (more for ribosome). Chosen by hand; affects representational capacity and runtime.
  • Density threshold for seeding Gaussians = not specified numerically (user-defined)
    Voxels with density below the threshold are discarded when seeding the coarse GMM (Supplement C). Manual choice affects whether flexible, low-density regions are covered.
  • Regularization weights lambda_z and lambda_f = lambda_z=0.1, lambda_f=0.01
    Hyperparameters for latent and feature-offset priors; selected without a reported tuning or cross-validation procedure.
  • Feature dimensionality E and latent dimensionality D = E=24, D=4
    Fixed for all experiments without sensitivity analysis.
  • Initial Gaussian scale and amplitude = m=0.15; s=0.02 (part discovery), s=0.01 (part-aware)
    Manual initial values; the paper notes optimization dynamics are sensitive to initialization of Gaussian parameters (Supplement C).
  • Number of spatial clusters after feature clustering = 5 (IgG-1D), 6 (IgG-RL)
    Feature clustering is followed by spatial k-means to split large parts; the number is chosen manually.
  • Training epochs = 15 or 50 (part discovery), 30 (part-aware)
    Chosen per dataset; no early-stopping or validation criterion reported.
assumptions (6)
  • domain assumption The 3D density of a biomolecule can be represented as a sum of isotropic 3D Gaussians, and the 2D projection is the analytic sum of 2D Gaussians (Eq. 1 and 2).
    This is the image formation model used throughout; it is standard in GMM cryo-EM methods, but it restricts the density representation.
  • domain assumption Per-particle poses (rotations and translations) and CTFs are known or estimated by external software (CryoSPARC, RELION).
    The method is not ab initio; it requires a rigid reconstruction and poses as input (Supplement C).
  • domain assumption Latent codes z follow a zero-mean Gaussian prior and feature offsets are small (Eq. 7).
    These priors regularize the model but impose a specific smooth structure on the learned heterogeneity space.
  • ad hoc to paper k-means++ clustering on learned Gaussian features yields a semantically meaningful part decomposition without direct supervision.
    The central part discovery claim rests on the empirical observation that features group by motion/occupancy coherence (Sec. 3.2); this is not guaranteed by the model or by theory.
  • domain assumption Per-Conformation FSC on CryoBench is a valid measure of heterogeneity reconstruction quality.
    The paper adopts the benchmark metric from CryoBench (Sec. 4); this metric assumes known ground-truth states and is used to rank methods.
  • standard math The rendering grid and separability tricks compute the projection accurately.
    Gaussian separability (Eq. 12) is exact for axis-aligned 2D Gaussians; no approximation beyond grid discretization.
invented entities (1)
  • Part anchors
    purpose: Define rigid-body reference frames for clusters of Gaussians, enabling hierarchical motion and occupancy modulation.
    Anchors are computational constructs (center and feature vector) introduced by the model. They have no independent physical evidence beyond the reconstruction quality; they are not new particles or forces, but they are invented abstractions that the central claim depends on.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery." pith.science (2026). https://pith.science/paper/DLPWSWLU

@misc{pith2026250609063,
  author       = {Pith},
  title        = {Pith review of: Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DLPWSWLU}},
  note         = {Machine review of arXiv:2506.09063}
}
read the original abstract

Cryo-EM is a transformational paradigm in molecular biology where computational methods are used to infer 3D molecular structure at atomic resolution from extremely noisy 2D electron microscope images. At the forefront of research is how to model the structure when the imaged particles exhibit non-rigid conformational flexibility and compositional variation where parts are sometimes missing. We introduce a novel 3D reconstruction framework with a hierarchical Gaussian mixture model, inspired in part by Gaussian Splatting for 4D scene reconstruction. In particular, the structure of the model is grounded in an initial process that infers a part-based segmentation of the particle, providing essential inductive bias in order to handle both conformational and compositional variability. The framework, called CryoSPIRE, is shown to reveal biologically meaningful structures on complex experimental datasets, and establishes a new state-of-the-art on CryoBench, a benchmark for cryo-EM heterogeneity methods.

Figures

Figures reproduced from arXiv: 2506.09063 by the authors.

Figure 1
Figure 1. (A) Based on a stack of noisy particle images, (B) CryoSPIRE learns a part-based Gaussian [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of CryoSPIRE. (A) To infer parts, we optimize a coarse GMM with neural networks that generate Gaussian amplitudes and centers, conditioned on image latent codes and Gaussian features. (B) Clustering on learned Gaussian features provides meaningful parts. The CryoSPIRE model comprises one anchor and a set of Gaussians per part. (C) MLPs generate the rigid-body motion of each anchor (top), per-Gaussian displa… view at source ↗
Figure 3
Figure 3. Per-Conformation FSC on CryoBench datasets. Error bars indicate standard deviation [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Results on IgG-1D [15]. (A) Due to large motion, the Fab domain (circled) is smeared out in rigid reconstruction, while our part discovery model identifies this domain and resolves its structure and motion, providing good initialization for subsequent modeling. (B) For…
Figure 5
Figure 5. Figure 5: Results on IgG-RL [15]. (A) The feature space, F, shows two parts (green and orange) corresponding to the flexible Fab domain and the rigid core. Subsequent spatial clustering yields six parts. (B) The latent space, Z, is colored with Fab orientation along with four sa…
Figure 6
Figure 6. Figure 6: Results on Ribosembly [15] (A) Gaussian feature space, F, showing eight major parts identified through clustering. (B) Heterogeneity latent space, Z, colored coded with the ground-truth compositional state. (C) Visualizations of 3D density maps corresponding to seven p…
Figure 7
Figure 7. Figure 7: Results on Large Ribosomal Subunit (EMPIAR-10076 [ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Results on Pre-Catalytic Spliceosome (EMPIAR-10180 [ [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Estimated motion of Gaussians for 30 states of IgG-1D. The baselines fail to cap￾ture local rigidity. the part-aware model with four anchors. To illustrate structural variability, we run PCA on the heterogeneity latent space, Z, and extract two principal directions ill…
Figure 10
Figure 10. Figure 10: Per-Image FSCs on CryoBench datasets. Error bars indicate standard deviation across [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Complete qualitative results on Ribosembly [ [PITH_FULL_IMAGE:figures/full_fig_p018_11.png]
Figure 12
Figure 12. Figure 12: Qualitative comparison of CryoSPIRE with four state-of-the-art methods on CryoBench [PITH_FULL_IMAGE:figures/full_fig_p019_12.png]
Figure 13
Figure 13. Figure 13: More qualitative result on Large Ribosomal Subunit (EMPIAR-10076 [ [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Qualitative comparison of CryoSPIRE with 3DVA [ [PITH_FULL_IMAGE:figures/full_fig_p021_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages

  1. [1]

    K-Means++: The advantages of careful seeding.Proc

    David Arthur and Sergei Vassilvitskii. K-Means++: The advantages of careful seeding.Proc. 18th Annual ACM-SIAM Symposium on Discrete Algorithms, page 1027–1035, 2007. 5, 15

  2. [2]

    Per- Gaussian embedding-based deformation for deformable 3d Gaussian splatting

    Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per- Gaussian embedding-based deformation for deformable 3d Gaussian splatting. InEuropean Conference on Computer Vision (ECCV), pages 321–335. Springer, 2024. 1, 3

  3. [3]

    Building proteins in a day: Efficient 3d molecular reconstruction.Proc IEEE Conf

    Marcus A Brubaker, Ali Punjani, and David J Fleet. Building proteins in a day: Efficient 3d molecular reconstruction.Proc IEEE Conf. Computer Vision and Pattern Recognition (CVPR),

  4. [4]

    Deep learning-based mixed-dimensional Gaussian mixture model for characterizing variability in cryo-EM.Nature Methods, 18(8):930–936, 2021

    Muyuan Chen and Steven J Ludtke. Deep learning-based mixed-dimensional Gaussian mixture model for characterizing variability in cryo-EM.Nature Methods, 18(8):930–936, 2021. 2, 3

  5. [5]

    Schmid, and Wah Chiu

    Muyuan Chen, Michael F. Schmid, and Wah Chiu. Improving resolution and resolvability of single particle cryoem structures using Gaussian mixture models.Nature Methods, 21:37–40,

  6. [6]

    Integrating molecular models into cryoem heterogeneity analysis using scalable high-resolution deep Gaussian mixture models.Journal of Molecular Biology, 435(9):168014, 2023

    Muyuan Chen, Bogdan Toader, and Lederman Roy. Integrating molecular models into cryoem heterogeneity analysis using scalable high-resolution deep Gaussian mixture models.Journal of Molecular Biology, 435(9):168014, 2023. 2, 3, 15

  7. [7]

    Modular assembly of the bacterial large ribosomal subunit.Cell, 167(6):1610–1622,

    Joseph H Davis, Yong Zi Tan, Bridget Carragher, Clinton S Potter, Dmitry Lyumkis, and James R Williamson. Modular assembly of the bacterial large ribosomal subunit.Cell, 167(6):1610–1622,

  8. [8]

    V olume rendering.ACM SIGGRAPH, 22(4):65–74, 1988

    Robert A Drebin, Loren Carpenter, and Pat Hanrahan. V olume rendering.ACM SIGGRAPH, 22(4):65–74, 1988. 3

Show all 49 references
  1. [9]

    Frank and L

    J. Frank and L. Al-Ali. Signal-to-noise ratio of electron micrographs obtained by cross correla- tion.Nature, 256(5516):376–379, 1975. 16 10

  2. [10]

    Cryo-EM heterogeneity analysis using regularized covari- ance estimation and kernel regression.Proc

    Marc Aurèle Gilles and Amit Singer. Cryo-EM heterogeneity analysis using regularized covari- ance estimation and kernel regression.Proc. National Academy of Science, 122(9):e2419140122,

  3. [11]

    Sugar: Surface-aligned Gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering

    Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned Gaussian splatting for efficient 3D mesh reconstruction and high-quality mesh rendering. InProc. CVPR, 2024. 3

  4. [12]

    Harauz and M

    G. Harauz and M. van Heel. Optimal determination of particle orientation, absolute hand, and contrast loss in single-particle electron cryomicroscopy.Optik, 73:146–156, 1986. 16

  5. [13]

    Estimating confor- mational landscapes from cryo-em particles by 3d zernike polynomials.Nature Communications, 14(1):154, 2023

    David Herreros, Roy R Lederman, James M Krieger, Amaya Jiménez-Moreno, Marta Martínez, David Myška, D Strelak, J Filipovic, Carlos OS Sorzano, and José M Carazo. Estimating confor- mational landscapes from cryo-em particles by 3d zernike polynomials.Nature Communications, 14(1...

  6. [14]

    High-resolution real-space reconstruction of cryo-em structures using a neural field network.Nature Machine Intelligence, 6:892–903, 2024

    Yue Huang, Chengguang Zhu, Xiaokang Yang, and Manhua Liu. High-resolution real-space reconstruction of cryo-em structures using a neural field network.Nature Machine Intelligence, 6:892–903, 2024. 2

  7. [15]

    Hanson, Pilar Cossio, and Ellen D

    Minkyu Jeon, Rishwanth Raghu, Miro Astore, Geoffrey Woollard, Ryan Feathers, Alkin Kaz, Sonya M. Hanson, Pilar Cossio, and Ellen D. Zhong. CryoBench: Diverse and challeng- ing datasets for the heterogeneity problem in cryo-em. InAdvances in Neural Information Processing System...

  8. [16]

    3D Gaus- sian splatting for real-time radiance field rendering.ACM Transactions on Graphics (Proc

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuehler, and George Drettakis. 3D Gaus- sian splatting for real-time radiance field rendering.ACM Transactions on Graphics (Proc. SIGGRAPH), 2023. 3, 15

  9. [17]

    3D Gaussian splatting as Markov chain Monte Carlo

    Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3D Gaussian splatting as Markov chain Monte Carlo. InProc. NeurIPS, 2024. 3

  10. [18]

    The Resolution Revolution: Advances in detector technology and image processing are yielding high-resolution electron cryo-microscopy structures of biomolecules

    Werner Kühlbrandt. The Resolution Revolution: Advances in detector technology and image processing are yielding high-resolution electron cryo-microscopy structures of biomolecules. Science, 343(6178):1443–1444, 2014. 1

  11. [19]

    Revealing biomolecular structure and motion with neural ab initio cryo-em reconstruction.bioRxiv, 2024

    Axel Levy, Michal Grzadkowski, Frederic Poitevin, Francesca Vallese, Oliver B Clarke, Gordon Wetzstein, and Ellen D Zhong. Revealing biomolecular structure and motion with neural ab initio cryo-em reconstruction.bioRxiv, 2024. 1, 2, 5, 6, 17, 19

  12. [20]

    Amortized inference for heterogeneous reconstruction in cryo-EM.Advances in Neural Information Processing Systems, 35:13038–13049, 2022

    Axel Levy, Gordon Wetzstein, Julien NP Martel, Frederic Poitevin, and Ellen Zhong. Amortized inference for heterogeneous reconstruction in cryo-EM.Advances in Neural Information Processing Systems, 35:13038–13049, 2022. 1

  13. [21]

    Scaffold- gs: Structured 3d Gaussians for view-adaptive rendering

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold- gs: Structured 3d Gaussians for view-adaptive rendering. InProc IEEE/CVF Conf Computer Vision and Pattern Recognition (CVPR), pages 20654–20664, 2024. 2, 3

  14. [22]

    Dynamic 3D Gaussians: Tracking by persistent dynamic view synthesis

    Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3D Gaussians: Tracking by persistent dynamic view synthesis. InProc. 3DV, 2024. 3

  15. [23]

    Optical models for direct volume rendering.IEEE Trans

    Nelson Max. Optical models for direct volume rendering.IEEE Trans. Vis. Comput. Graph., 1(2):99–108, 2002. 3

  16. [24]

    NERF: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoor- thi, and Ren Ng. NERF: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99–106, 2022. 2

  17. [25]

    Multi-body refinement of cryo-em images in relion

    Takanori Nakane and Sjors HW Scheres. Multi-body refinement of cryo-em images in relion. InCryoEM: Methods and Protocols, pages 145–160. Springer, 2020. 3

  18. [26]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. InProc IEEE/CVF Conf Computer Vision and Pattern Recognition (CVPR), pages 165–174, 2019. 5 11

  19. [27]

    Nerfies: Deformable neural radiance fields

    Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. InProc IEEE/CVF International Conference on Computer Vision (CVPR), pages 5865–5874, 2021. 5

  20. [28]

    Pawel A. Penczek. Three-dimensional spectral signal-to-noise ratio for a class of reconstruction algorithms.Journal of Structural Biology, 138(1):34–46, 2002. 16

  21. [29]

    Ucsf chimerax: Structure visualization for researchers, educators, and developers.Protein science, 30(1):70–82, 2021

    Eric F Pettersen, Thomas D Goddard, Conrad C Huang, Elaine C Meng, Gregory S Couch, Tristan I Croll, John H Morris, and Thomas E Ferrin. Ucsf chimerax: Structure visualization for researchers, educators, and developers.Protein science, 30(1):70–82, 2021. 14

  22. [30]

    Structure of a pre-catalytic spliceosome

    Clemens Plaschka, Pei-Chun Lin, and Kiyoshi Nagai. Structure of a pre-catalytic spliceosome. Nature, 546(7660):617–621, 2017. 5, 8, 9, 15, 20, 21

  23. [31]

    D-nerf: Neural radiance fields for dynamic scenes

    Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. InProc IEEE/CVF Conf on Computer Vision and Pattern Recognition (CVPR), pages 10318–10327, 2021. 5

  24. [32]

    3D Variability Analysis: Directly resolving continuous flexibility and discreteheterogeneity from single particle cryo-EM images.Journal of Structural Biology, 213:107702, 2021

    Ali Punjani and David J Fleet. 3D Variability Analysis: Directly resolving continuous flexibility and discreteheterogeneity from single particle cryo-EM images.Journal of Structural Biology, 213:107702, 2021. 1, 2, 5, 6, 17, 19, 20, 21

  25. [33]

    3DFlex: Determining structure and motion of flexible proteins from cryo-em.Nature Methods, 20:860–870, 2023

    Ali Punjani and David J Fleet. 3DFlex: Determining structure and motion of flexible proteins from cryo-em.Nature Methods, 20:860–870, 2023. 1, 2, 3, 5, 6, 17, 19, 20, 21

  26. [34]

    CryoSPARC: Algorithms for rapid unsupervised cryo-em structure determination.Nature Methods, 14:290–296, 2017

    Ali Punjani, John L Rubinstein, David J Fleet, and Marcus A Brubaker. CryoSPARC: Algorithms for rapid unsupervised cryo-em structure determination.Nature Methods, 14:290–296, 2017. 15, 20

  27. [35]

    Cryo-em captures early ribosome assembly in action.Nature Communications, 14(1):898, 2023

    Bo Qin, Simon M Lauer, Annika Balke, Carlos H Vieira-Vieira, Jörg Bürger, Thorsten Mielke, Matthias Selbach, Patrick Scheerer, Christian MT Spahn, and Rainer Nikolay. Cryo-em captures early ribosome assembly in action.Nature Communications, 14(1):898, 2023. 5

  28. [36]

    P. B. Rosenthal and R. Henderson. Optimal determination of particle orientation, absolute hand, and contrast loss in single-particle electron cryomicroscopy.J. Molecular Biology, 333(4):721–745, 2003. 6, 16

  29. [37]

    Sjors H. W. Scheres. RELION: Implementation of a Bayesian approach to cryo-em structure determination.Journal of Structural Biology, 180(3):519 – 530, 2012. 15

  30. [38]

    Processing of structurally heterogeneous cryo-em data in RELION

    Sjors H W Scheres. Processing of structurally heterogeneous cryo-em data in RELION. In R A Crowther, editor,The Resolution Revolution: Recent Advances In cryoEM, volume 579 of Methods in Enzymology, pages 125–157. Academic Press, 2016. 1

  31. [39]

    Disentangling conformational states of macromolecules in 3D-EM through likelihood optimization.Nature Methods, 4(1):27–29, 2007

    Sjors H W Scheres, Haixiao Gao, Mikel Valle, Gabor T Herman, Paul P B Eggermont, Joachim Frank, and Jose-Maria Carazo. Disentangling conformational states of macromolecules in 3D-EM through likelihood optimization.Nature Methods, 4(1):27–29, 2007. 6, 17

  32. [40]

    Johannes Schwab, Dari Kimanius, Alister Burt, Tom Dendooven, and Sjors H. W. Scheres. DynaMight: Estimating molecular motions with improved reconstruction from cryo-em images. Nature Methods, 21:1855–1862, 2024. 2, 3, 5, 9

  33. [41]

    Computational methods for single-particle electron cryomi- croscopy.Annual review of biomedical data science, 3(1):163–190, 2020

    Amit Singer and Fred J Sigworth. Computational methods for single-particle electron cryomi- croscopy.Annual review of biomedical data science, 3(1):163–190, 2020. 14

  34. [42]

    Unser, B

    M. Unser, B. L. Trus, and A. C. Steven. A new resolution criterion based on spectral signal-to- noise ratios.Ultramicroscopy, 23(1):39–51, 1987. 16

  35. [43]

    Fourier shell correlation threshold criteria.Journal of structural biology, 151(3):250–262, 2005

    Marin Van Heel and Michael Schatz. Fourier shell correlation threshold criteria.Journal of structural biology, 151(3):250–262, 2005. 6, 16

  36. [44]

    4D Gaussian splatting for real-time dynamic scene rendering

    Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4D Gaussian splatting for real-time dynamic scene rendering. In Proc. CVPR, 2024. 3 12

  37. [45]

    Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction

    Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3D Gaussians for high-fidelity monocular dynamic scene reconstruction. InProc. CVPR, 2024. 3

  38. [46]

    Mip-splatting: Alias-free 3D Gaussian splatting

    Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3D Gaussian splatting. InProc. CVPR, 2024. 3

  39. [47]

    Zhong, Tristan Bepler, Bonnie Berger, and Joseph H

    Ellen D. Zhong, Tristan Bepler, Bonnie Berger, and Joseph H. Davis. CryoDRGN: Reconstruc- tion of heterogeneous cryo-em structures using neural networks.Nature Methods, 18:176–185,

  40. [48]

    gold standard

    Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting.IEEE Transactions on Visualization and Computer Graphics, 8(3):223–238, 2002. 3 13 Supplementary Material Reconstructing Heterogeneous Biomolecules via Hierarchical Gaussian Mixtures and Part...

  41. [2021]

    1, 2, 5, 6, 17, 19, 20, 21

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.