acceptodds
Under review as a conference paper at ICLR 2027

Atomic modeling of protein dynamics from cryo-EM with co-folding priors

Abstract

Understanding how proteins and other biomacromolecular machines function requires characterizing not only individual structures, but their full conformational landscapes. New protein foundation models provide powerful priors over atomic structure, and recent work has leveraged experimental measurements to guide the generation of individual structures at inference time. However, recovering atomic conformational landscapes from experimental data remains an open problem. We introduce CryoCHEF, a method for all-atom modeling of protein onformational eterogeneous nsembles using co-olding priors. CryoCHEF proposes a new, general framework for mapping learned conformational representations from experimental cryo-EM data into the latent space of a pretrained co-folding model. Specifically, CryoCHEF learns a conformation-conditioned neural field over the pair representations of a pretrained co-folding model, inducing a continuous atomic ensemble conditioned on experimental data. Across synthetic benchmarks and experimental datasets, CryoCHEF recovers continuous structural ensembles with side-chain-level detail, captures known global and local conformational motions, and enables efficient generation of atomic trajectories across experimentally observed conformational landscapes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryo-EM, protein folding, generative models

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_9z8DOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes CryoCHEF, a framework for mapping heterogeneous cryo-EM data into the latent space of a trained protein generator model in order to get diverse ensembles.

Critical strengths

Useful framing and the design is light and efficient.

Critical weaknesses

See questions

Questions for authors
  1. Could the authors add comparisons to baselines on the real dataset?
  2. Could the authors run one or more of the methods that guide based on a single density map, for at least one synthetic and one experimental dataset?
  3. If I understand the method correctly the weights need to be trained for every instance of protein separately, how long does it take to train it? How does it scale with sequence length? How does it compare to the methods from question (2)?
  4. Could the authors add an ablation over the number of training latents?
  5. Does CryoCHEF recover rare yet physical conformations?
Reviewer_xcRuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents CryoCHEF, a method for mapping conformational latent spaces learned from heterogeneous cryo-EM data into continuous all-atom structural ensembles. CryoCHEF learns a latent-conditioned bias over the pair representation of a frozen ESMFold2 model, supervised by map-to-model agreement through differentiable density rendering. The method operates downstream of the heterogeneous reconstruction approaches such as cryoDRGN and DynaMight, enabling efficient structure generation across their learned conformational landscapes.

Critical strengths
  • CryoCHEF achieves strong atomic modeling results across multiple systems. More importantly, the authors demonstrate that CryoCHEF works well on experimental datasets rather than only on synthetic examples.

  • Compared with per-map guidance approaches, CryoCHEF amortizes atomic model building over a pre-existing continuous cryo-EM latent space, enabling atomic structures to be queried throughout the reconstructed conformational landscape.

Critical weaknesses
  • CryoCHEF is closely related to CryoSampler, which also combines a pretrained folding prior, heterogeneous cryo-EM map ensembles, differentiable density rendering to recover atomic conformational ensembles. The main difference is that CryoCHEF explicitly learns a mapping from a precomputed reconstruction latent space to ESMFold2 pair representations, whereas CryoSampler learns its own deformation distribution around a Boltz-2 reference. While preserving correspondence with the upstream latent space is potentially useful, the paper does not clearly demonstrate its advantage in accuracy, efficiency, or scientific interpretation.

  • The claims regarding local flexibility and “side-chain-level detail” (abstract, L23-24) are not sufficiently supported by the experiments. For the real datasets, training is restricted to 10 Å for spike and integrin and 15 Å for HER2, where the density cannot directly constrain side-chain conformations. Based on the presented results, these local variations are not supported by the experimental density and are likely introduced by the ESMFold2 prior. Therefore, CryoCHEF’s advantage over the baselines is demonstrated mainly on synthetic datasets, while the qualitative results on real datasets do not clearly show what additional information the method provides.

  • The related-work discussion omits CryoSampler and CryoACE, both of which are closely related to the proposed setting and should be cited and discussed.

Questions for authors
  • What motivated the choice of ESMFold2 as the folding prior instead of more widely used models such as Boltz or Protenix?
Reviewer_iYufOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes CryoCHEF, which learns all-atom protein conformation ensemble from cryo-EM images. By combining the latent representation of a cryo-EM reconstruction model into the pair representation of a protein structure prediction model CryoCHEF recovers accurate all-atom representations. Moreover, in case studies including synthetic and real cryo-EM data, CryoCHEF is able to model the motions of the proteins by interpolating the principal components of the latent space. Experiments show improved structure reconstruction accuracy and distribution accuracy over existing reconstruction methods.

Critical strengths
  1. CryoCHEF mitigate the gap between the 3D density of existing reconstruction and the all-atom or coarse-grained conformation ensemble of proteins.
  2. CryoCHEF shows advanced structure and distributional accuracy on synthetic datasets and recovers motion trajectories over an interpolated trajectory of latent space.
  3. The proposed method is simple and general. It is possible to adapt CryoCHEF to different combinations of reconstruction models and structure prediction models with pair feature.
  4. The paper is clearly written and well demonstrated.
Critical weaknesses
  1. A cryo-EM image can implicitly correspond to multiple all-atom protein conformations. Enforcing the model to recover a single all-atom conformation from each cryo-EM image is not a well-defined problem. Thus, it is hard to evaluate the fine-grained structure quality or distributional quality of generated all-atom conformations, especially for the experimental data. In the paper, the recover quality is only demonstrated in synthetic data.
  2. The 3D density can have difficulty in reflecting the details of loop structures, but may already contain the dominant conformational heterogeneity. For example, the spike RBD motion of SARS-Cov-2 virus is already captured by CryoDRGN in CryoBench[1]. In that case, it is unclear what atomic modeling adds to the existing reconstruction methods.
  3. Some related work is not discussed. For example, CryoSampler[2] also recovers all-atom models from map ensemble by cryo-EM. It would be valuable to discuss the difference and advantage of CryoCHEF over this related work.

Reference: [1] CryoBench: Diverse and Challenging Datasets for the Heterogeneity Problem in Cryo-EM [2] Modeling Atomic Conformational Ensembles of Proteins via Test-Time Supervision of Boltz-2 on Cryo-EM Density Maps

Questions for authors
  1. For the synthetic data, how the cryo-EM image is generated from MD data?
  2. How to demonstrate that atomic level recovering is more accurate than density-only reconstruction, in this paper, the CryoDRGN itself?
  3. Please check the weakness part for other questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.