acceptodds
Withdrawn from ICLR 2027

CryoSampler: Amortized Atomic Model Building Across Cryo-EM Conformational Ensembles

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Cryo-electron microscopy (cryo-EM) increasingly resolves proteins in motion, recovering an ensemble of maps spanning a molecule's conformational landscape. Converting these maps into atomic coordinates is an inverse problem known as model building, a necessary step for every structure deposited in the Protein Data Bank, which in turn supplies the training data for the next generation of AlphaFold-like models. However, existing model building methods treat each map as an independent inverse problem, ignoring that every map in an ensemble is an observation of the same molecule. We ask whether solving this inverse problem jointly across the ensemble with a single amortized model yields better atomic models than per-map inference. In this work, we propose CryoSampler, an amortized framework for ensemble model building that trains a 3D variational autoencoder to map each cryo-EM map to a neural deformation field over a shared reference structure from a frozen Boltz-2 model, supervised end-to-end through a differentiable cryo-EM forward model and stereochemical restraints. We find that amortization acts as an implicit structural prior that produces atomic models of higher geometric quality at lower compute cost, achieving state-of-the-art performance across several experimental datasets. These results establish amortized inference across conformational ensembles as a simple and effective principle for converting heterogeneous cryo-EM maps into high-quality atomic structures.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
cryo-electron microscopy, cryo-EM, structural biology, conformations, ensembles, protein structure

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_oZdHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CryoSampler targets atomic model building from ensembles of reconstructed cryo-EM density maps, rather than heterogeneous reconstruction directly from particle images. The method uses a 3D VAE to predict neural deformation fields over a frozen Boltz-2 reference structure, and trains the model with differentiable density rendering together with stereochemical and backbone restraints. The key contribution is a shared parameterization across conformational states: it keeps atom correspondence consistent while balancing density fit against structural validity.

Critical strengths
  1. Strong evidence of improved geometric regularization. Table 2 provides the clearest positive result: at essentially identical CC_mask values of 0.76, CryoSampler reduces clashscore from 168.7 to 19.9 and rotamer outliers from 3.97% to 0.19%, using the same losses and reference initialization as direct optimization. Maintaining density fit while substantially improving geometry is a meaningful achievement; the absence of a CC improvement does not diminish this particular result. However, it establishes better geometric regularization, not necessarily more accurate conformational recovery.

  2. A practically relevant downstream task. Converting heterogeneous density maps into corresponding full-atom models is a useful objective. Operating at the reconstructed-map level also makes the framework compatible, in principle, with different upstream reconstruction methods. This is a meaningful distinction from approaches that integrate atomic modeling into a particular particle-level reconstruction pipeline.

Critical weaknesses
  1. The connection to experimentally supported conformational changes remains insufficiently established. My main concern is the gap between producing plausible atomic models and recovering accurate conformational differences. Density maps provide experimental supervision, but good density agreement and favorable stereochemistry do not, by themselves, establish that the inferred state-dependent atomic displacements are correct. The manuscript explicitly acknowledges that ground-truth structures are unavailable and defines its performance claims through map–model fit and geometric validation. These are appropriate model-building criteria, but they do not directly validate conformational accuracy. Consequently, the evidence supports a useful fitting and regularization method more directly than it supports the broader motivation of recovering experimentally grounded atomic ensembles for studying conformational variability. A controlled benchmark with known atomic conformations would help establish this additional scientific value.

  2. The contribution of neural parameterization is not sufficiently separated from cross-map amortization. Table 2 changes both the coordinate representation and the sharing of parameters across maps. It therefore does not isolate which aspect produces the large geometric improvements. Table 3 provides a more targeted comparison: joint training slightly reduces density fit on all three tested systems, improves MolProbity on two, and provides no geometric improvement on integrin. The authors acknowledge this trade-off. These results are consistent with additional regularization from parameter sharing. They do not yet establish that joint inference recovers conformational information that cannot be recovered by fitting each map separately. The distinction matters because the benefit of a constrained neural representation is not equivalent to the benefit of learning jointly across related observations.

  3. The distinctive machine-learning contribution remains unclear. The paper’s own related-work discussion identifies precedents for reference-based VAE deformation and for neural parameterizations acting as implicit priors. Its novelty therefore appears primarily in the particular ensemble model-building formulation and system integration. That can constitute a worthwhile contribution, but the broader methodological claim would be stronger if the experiments isolated what is gained specifically from amortization, rather than from reference preservation or constrained deformation. The reported absence of a measurable effect from removing the KL term also leaves the necessity of the variational formulation unclear.

Questions for authors

Q1. Could the authors clarify the motivation and empirical benefit of the encoder–latent–decoder design, particularly its variational component? Section 5.3 reports that removing the KL term has no measurable effect on model-building metrics, which leaves the benefit of the VAE formulation unclear. A comparison with a deterministic encoder–decoder would help establish the contribution of the variational component, while a shared deformation decoder with directly optimized per-map latent codes would clarify the contribution of encoder-based amortization. Do these simpler alternatives achieve comparable performance?

Q2. What demonstrates useful information sharing across maps? Does increasing the number of related maps improve recovery of a noisy or held-out state from the same molecular system? Such an experiment could test whether joint training transfers useful conformational information rather than simply imposing stronger regularization.

Q3. How are the individual loss terms normalized, and how were their weights selected? Although using the same weights across systems is encouraging, it remains unclear how sensitive the results are to these choices.

Q4. Can the method recover known conformational changes? Could you provide simulated-data experiments with ground-truth atomic conformations, measuring state-specific coordinate errors and errors in relative motions across states?

Reviewer_K2xAOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CryoSampler, a method for fitting atomic models to ensembles of cryo-EM density maps. Given a shared reference structure, a spatial VAE predicts map-conditioned atomic displacements. Training combines a differentiable density-rendering objective with structural restraints. The paper evaluates model building on four protein systems and additionally explores transferring conformational variation to related proteins through a latent generative model.

Critical strengths
  1. The neural parameterization provides useful regularization for a relevant model-building problem. The comparison against direct coordinate optimization with the same initialization and losses is compelling: Table 2 reports comparable map correlation but substantially fewer atomic clashes. This supports the practical value of the proposed parameterization for producing structurally plausible atomic models.

  2. The evaluation includes several informative controls and useful supplementary material. The authors compare joint and per-map training, ablate structural and backbone restraints, and examine sensitivity to preprocessing. Reporting EMRinger and CaBLAM alongside map correlation and directly restrained geometry metrics provides a broader assessment. The supplementary package includes training code, configurations, density maps, and reference structures.

Critical weaknesses
  1. The specific benefit of joint ensemble modeling remains insufficiently established. Table 2 demonstrates the benefit of neural parameterization over direct coordinate optimization, but does not isolate the benefit of learning across conformations. In the more direct comparison in Table 3, per-map training achieves higher map correlation on all three systems. Joint training improves aggregate geometry on two systems, provides essentially no such improvement on integrin, and offers a modest runtime reduction. These results establish a useful trade-off, but do not clearly show that the model learns shared conformational information beyond imposing stronger regularization through parameter sharing. A comparison against a suitably regularized per-map model, ideally across a fit–geometry trade-off curve, would help clarify this distinction.

  2. The evidence for recovering accurate conformational changes is less convincing than the evidence for maintaining plausible geometry. The method starts from a predicted reference structure and explicitly restrains its deformation. Consequently, favorable stereochemical metrics can partly reflect preservation of the reference rather than accurate recovery of map-specific conformations. The additional validation metrics are useful, but I would like stronger evidence in regions undergoing substantial conformational change, sensitivity to alternative initial conformations, or validation against independent half-maps. I did not find an explicit independent half-map validation experiment in the supplied material. This concern is compounded by dependence on preprocessing: some datasets use reference trimming or masks informed by deposited structures, and the substantial degradation on integrin in the preprocessing ablation suggests that these choices materially affect performance.

Overall, the paper presents a useful approach with meaningful implementation work and several informative experiments. My main reservation is that the strongest evidence supports neural parameterization as a regularizer, while the distinctive benefits of joint conformational modeling and the accuracy of the recovered structural changes remain less firmly established. I currently lean toward weak rejection, but additional evidence isolating the benefit of joint training and independently validating conformational accuracy could change my assessment.

Questions for authors
  1. Can a per-map model with appropriately tuned regularization match the joint model’s geometry at a comparable level of map fit? What evidence specifically demonstrates useful information sharing across conformations?

  2. How sensitive are the recovered conformational changes to the reference structure? Do different plausible initial conformations converge to consistent results in regions supported by the experimental density? Please also clarify the prior structural information required for preprocessing when deposited structures are unavailable.

  3. Is independent half-map validation available, or another analysis that distinguishes improved recovery of experimental signal from retention of the reference geometry?

  4. The transfer experiments cover only two targets within the TRP family. Table 8 shows slightly higher precision than AF-Multimer-Cluster, but lower recall and higher Wasserstein distance on both targets. How should these results delimit the claim about transferable conformational modeling?

  5. Could the authors provide the configurations, data specification, and evaluation scripts corresponding to the updated transfer results in Table 8? The supplied transfer configuration uses four maps from a single TRPV3 principal component and targets TRPV5, whereas the updated experiment uses a broader training protocol and evaluates both TRPV1 and TRPV5. This observation is based on static inspection; the training code was not executed.

Reviewer_LfqDOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

CryoSampler is an approach to assign an ensemble of atomic models to a set of cryoEM maps of a complex in multiple conformational states. It requires a Boltz-2 atomic model and it learns displacements of the atoms through a shared neural network to match the different conformational states.

Critical strengths

The approach assigns an ensemble of atomic models to a set of cryoEM maps of the same complex in various conformational states rather than just a single model for a single state. It requires an initial atomic model and it learns displacements of the atoms through a shared neural network to match the different conformational states. The joint optimizaiton should be better suited for the applicaiotn.

Critical weaknesses

Novelty is limited: The authors fail to highlight atomic model fitting capabilities of previous work for heterogeneous data. Amongst other methods, DynaMight can take atomic models as input. This is discussed in the original paper. Moreover, methods like DynaMight work directly with the 2D projections and jointly perform heterogeneity analysis as well as atomic model displacement, which retains far more information about the data rather than using a few 3D volumes generated by another software.

Evaluation is limited: The authors present very few experiments and they contain a limited number of volumes and heterogeneous states. Moreover, it’s not clear if the joint fitting provides an advantage over fitting each map separately. Additionally, the recovered atomic models are not evaluated against a ground truth.

It’s unclear why the Boltz-2 model takes such a central role in the paper, when the initial atomic model can be from any source, including the PDB database.

Questions for authors
  1. Can the authors clarify the novelty compared to methods like DynaMight?
  2. Are there any advantages compared to previous methods like DynaMight?
  3. Can the authors evaluate the method on larger, more diverse and more heterogeneous dataset?
  4. How does the approach handle more volume in eache ensemple?
  5. Can the authors show that the joint fitting has an advantage?
  6. Can the authors provide some quantitative evaluation of the generated atomic models?
  7. Can the authors clarify why Boltz-2 is necessary for their approach, rather than using an initial atomic model from any other source, such as the PDB?

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

Related papers

Discussion (0)

Sign in to comment.