acceptodds
Withdrawn from ICLR 2027

CryoHyper: Toward Reliable Conformational Spaces in Cryo-EM and Cryo-ET

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Resolving conformational heterogeneity is essential for understanding the functional mechanisms of macromolecular complexes. Existing continuous heterogeneity methods can generate smoothly varying structures but do not necessarily organize them according to structural relationships. As a result, structurally dissimilar conformations may become neighbors in the learned representation, while distinct motions can become entangled, complicating recovery of continuous conformational changes. This problem is compounded by noise and missing-wedge effects in cryo-electron tomography (cryo-ET), hindering atomic conformational analysis. We present CryoHyper, a framework for learning reliable conformational spaces from cryo-electron microscopy (cryo-EM) and cryo-ET. Guided by a reference atomic model, CryoHyper models domain motions in a hyperbolic subspace and local deformations in a Euclidean subspace. The learned conformational coordinates condition density generation, placing atomic models and density maps in the same conformational coordinate system. For cryo-ET, the framework combines learnable spectral weighting with reconstruction training using random rotations and artificial wedge masking to reduce the influence of noise and missing-wedge effects on conformational inference. Experiments on simulated and public datasets show that CryoHyper recovers physically plausible conformational changes and organizes them in closer agreement with the underlying structural relationships.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cryo-electron tomography, Cryo-electron microscopy, Conformational heterogeneity analysis, Hyperbolic representation learning, Variational autoencoding

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_EcDHOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CryoHyper is a method for recovering continuous conformational heterogeneity from cryo-EM and cryo-ET datasets. It decomposes the deformation of a reference atomic structure into global and local components, predicted from learned hyperbolic and euclidean latent spaces. The method is validated on a synthetic and two real datasets, demonstrating good conformation space organization and volume reconstruction.

Critical strengths

The recovered latent spaces have a nice continuous structure (e.g. Figure 2b) compared to the baselines, which could be helpful for interpretation in biological workflows. This is supported by improved information imbalance and neighborhood similarity metrics. Additionally, the method is demonstrated to recover heterogeneity from two real datasets (Figures 4 and 5), showing its applicability to experimental settings.

Critical weaknesses

The single-particle benchmark could be better designed. In particular, AD-FSC, BP-FSC, Map-CC, and SSIM all compare method-derived outputs to each other (e.g. structure vs map for AD-FSC), and are thus measures of self-consistency rather than correctness. These numbers cannot really be compared against the baselines. A compelling evaluation would include a standard synthetic data benchmark such as CryoBench, which is partially tested in the supplement. Reconstructed maps should be compared to GT maps in addition to latent space metrics. IgG-1D/IgG-RL could be useful to demonstrate on manifolds which do not expect to show U-shaped manifolds as in the 3 main datasets.

The motivation in the intro (e.g. "continuous paths should follow ordered structural changes") feels too strong. In particular, the demonstrated PC1 trajectories only capture directions of variation, not biophysically meaningful paths (which might be inferred by tracing low energy pathways through an estimated free energy landscape). And latent distances may be more correct in a ranking sense but the nonlinear encoding process can still distort distances if I understand the method correctly. Accordingly, the main contribution here is that a hyperbolically structured latent space can improve latent inference.

Questions for authors
  • Can you clarify the claim of "making latent distances more faithful to structural relationships between conformations"?
  • In Figure 3A, it appears that CryoHyper performs poorly relative to the baselines in a very local neighborhood (<5%) despite its stronger performance for large radii. Can you explain this?
  • The UMAPs in Figures 2,4, and 5 all have the same general shape. Is this a coincidental property of the data, or a bias of CryoHyper? Should we expect the method to work for more complex continuous manifolds?
Reviewer_HymuOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors present CryoHyper, a framework for learning conformational representations from cryo-EM particle images and cryo-ET subtomograms. The method combines a hyperbolic latent subspace for domain-level motions with a Euclidean subspace for local atomic deformations, using a reference atomic model and structural regularization. It further introduces an input-adaptation scheme for cryo-ET based on learnable spectral weighting and artificial missing-wedge augmentation. Experiments on simulated and experimental datasets show that CryoHyper produces coherent conformational spaces and particle subsets that support downstream refinement.

Critical strengths
  • The reported results are generally strong. CryoHyper achieves improved performance across several evaluation criteria compared to multiple baselines. It is also evaluated on multiple simulated and experimental datasets.

  • The method handles both single-particle cryo-EM and cryo-ET data.

Critical weaknesses

My main concern is the limited methodological novelty. The central new idea seems to be the use of a hyperbolic latent representation for domain-level motion, together with a Euclidean latent representation for local deformation. Most other components are closely related to existing methods: atomic-model-guided reconstruction and structural regularization are also largely inherited from cryoSTAR; domain-level rigid motion is related to prior segment-based approaches (e.g. multibody refinement, cryoSPHERE); and rotation/artificial-wedge augmentation has previously been explored in cryo-ET restoration. The explicit DSSP constraints, ECOD domain assignments, spectral weighting are useful additions, but they appear primarily to be engineering choices or combinations of existing ideas rather than individually substantial methodological contributions.

Besides the limited novelty, this also makes it difficult to determine which component is responsible for the improvements compared to the baselines. Comparisons against existing methods therefore establish the effectiveness of the complete system, but do not fully isolate the contribution of the proposed hyperbolic representation. The H4E4-E4E4 comparison is a useful control, but the ablation is evaluated only on EMPIAR-10073 using AD-FSC and refinement resolution, rather than metrics that directly assess whether latent distances preserve structural relationships. It would be better if the authors could compare H4E4 and E4E4 on simulated datasets using BNS and bidirectional Information Imbalance.

Questions for authors

Could the authors provide more density map comparisons with the baselines? Especially, for other methods that also use atomic models as constrains (e.g. cryoSTAR, cryoSPHERE), CryoHyper improves R-FSC and BP-FSC by approximately 0.5-2 Å, and matched visual comparisons would help demonstrate the practical significance of these improvements.

Reviewer_Jff7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CryoHyper, a new framework that aims to learn a reliable conformational latent space in both the cryo-EM and cryo-ET settings. Given a reference atomic model, the framework takes either cryo-EM particle images or cryo-ET subtomograms as input and extracts conformational embeddings with a CNN or a 3D CNN encoder, respectively. The key contribution is to represent protein motion with a joint hyperbolic and Euclidean representation, which captures the domain-level and local motions of the reference atomic model that align with the observations. The experiments show that CryoHyper can recover continuous domain and local motions of the atomic model in practice.

Critical strengths
  1. CryoHyper offers a unified, atomic-model-guided framework for both cryo-EM and cryo-ET, in which atomic models and density maps share the same latent coordinates. On EMPIAR-10073, particle subsets selected from its latent space refine to 4.57–5.73 Å and resolve head and foot motions that 3D classification misses.

  2. On the simulated cryo-ET benchmark, CryoHyper is clearly more robust to noise. At SNR 0.01 its BNS-AUC$_{20}$ is 41.88, compared with 24.35 for cryoDRGN-ET and 22.09 for OPUS-Tomo.

Critical weaknesses
  1. The motivation for using a hyperbolic subspace for domain motion is unclear. The paper only notes that hyperbolic distances grow toward the boundary; it does not explain why domain motions in cryo-EM should be hierarchical or tree-like. Rigid-body rotations live on SO(3), which is positively curved, and the different scales of domain and local motion could be handled by Euclidean latents with different variances. I would ask the authors to explain which property of cryo-EM conformational landscapes motivates a negatively curved latent space, and what concrete advantage it offers over a Euclidean latent such as CryoSTAR’s.

  2. The improvements on experimental data are modest, and the validity of the decoded atomic models is not evaluated. On EMPIAR-10073, the refinement resolution is essentially tied with cryoDRGN (5.02 vs 5.04 Å), BP-FSC trails RECOVAR and cryoDRGN, and all results come from a single training run. The paper also claims physically plausible conformational changes but reports no geometric validation of the intermediate atomic models. Missing checks include bond-length and angle deviations, clash scores, and Ramachandran statistics, particularly at domain hinges where rigid-body transforms meet. Even on the simulated 3NVQ and Spike-MD data, where ground-truth atomic models are available, the RMSD between decoded and ground-truth structures is not reported. I would ask the authors to provide more comprehensive evaluations regarding the biochemical properties of the deformed atomic model.

Questions for authors

please see my weakness section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.