acceptodds
Under review as a conference paper at ICLR 2027

HeteroGauss: Heterogeneity-Aware Gaussian Representation for Cryo-EM Reconstruction

17 views

Abstract

Single-particle cryo-EM captures biomolecular structures with continuous conformational and discrete compositional variability, yet each particle image is an extremely noisy 2D projection of one unknown state. Recovering heterogeneous ensembles requires sharing information across particles while preserving structural correspondence, allowing molecular elements to move, appear or disappear, and adapt spatial capacity to unresolved heterogeneous regions. Existing voxel, implicit-field, and deformation-based models typically satisfy only part of these requirements, either lacking explicit cross-state correspondence, relying on an established reference, or retaining fixed spatial support throughout reconstruction. We propose HeteroGauss, a shared, state-conditioned anisotropic Gaussian representation for heterogeneous cryo-EM reconstruction. Mechanistically, each Gaussian preserves identity across molecular states while its position, scale, orientation, and amplitude vary with the inferred latent state, enabling explicit modeling of both conformational and compositional variability. Learnable spatial anchors coordinate local deformation, while a heterogeneity-aware adaptive topology mechanism reallocates primitives using reconstruction gradients and state-dependent motion. Technically, a batched differentiable CUDA renderer enables efficient optimization of particle-specific Gaussian structures. Experiments on CryoBench and EMPIAR-10180 show that HeteroGauss reconstructs continuous and compositional heterogeneity with strong accuracy and retains substantially more reproducible high-frequency detail on experimental data. Codes are attached for evaluation. Project Webpage.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cryo-EM, 3D Gaussian Representation, Heterogeneous Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_qLTiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes HeteroGauss, a heterogeneous cryo-EM reconstruction method built on a shared set of anisotropic Gaussians conditioned on particle state. Each particle is assigned an inferred latent state, and the position, scale, orientation, and amplitude of every Gaussian depend on that state. This allows a single representation to capture both conformational and compositional heterogeneity. Learnable spatial anchors coordinate local deformations. An adaptive topology mechanism clones, splits, and prunes Gaussians according to reconstruction gradients and state-dependent motion. The authors also implement a batched differentiable CUDA renderer for efficiency.

Experiments cover three CryoBench datasets (IgG-1D, IgG-RL, and Ribosembly) and one experimental dataset (EMPIAR-10180). HeteroGauss reports the highest mean AU-FSC across the three synthetic benchmarks, and the authors show sharper high-frequency detail on the experimental data. Ablations examine anchor-based deformation, adaptive topology, amplitude modulation, and frequency-weighted reconstruction.

Critical strengths
  1. A flexible Gaussian representation that handles both kinds of heterogeneity.

HeteroGauss keeps a shared set of Gaussians with fixed correspondence across states, while letting their geometry and amplitude change with the latent state. This lets one representation cover both conformational and compositional heterogeneity. Learnable anchors allow local deformations, and adaptive topology adds capacity to regions that are hard to model. Overall, this is a sensible direction for heterogeneous cryo-EM reconstruction.

  1. Strong benchmark results, supported by ablations.

HeteroGauss obtains the highest mean AU-FSC on all three CryoBench datasets, with clear margins on IgG-1D and IgG-RL. On EMPIAR-10180, it shows better high-frequency agreement than RECOVAR. The ablations indicate that several components improve reconstruction quality, and adaptive topology gives the most consistent gains.

Critical weaknesses
  1. No comparison with the closest Gaussian-based heterogeneous reconstruction methods.

The core of HeteroGauss is a shared Gaussian representation whose geometry and amplitude depend on the latent state. Yet the main quantitative comparisons leave out the most closely related methods, notably CryoSPIRE and e2gmm, both of which use Gaussian representations for heterogeneous reconstruction.

CryoSPIRE is the most important omission. It also uses state-dependent Gaussians coordinated by anchors. The authors distinguish HeteroGauss from it by pointing to anisotropic covariances and adaptive spatial capacity, but provide no experiment showing that these differences actually lead to better results.

The comparisons with CryoDRGN, RECOVAR, 3DFlex, and other established methods are useful, but they do not show what HeteroGauss adds over its nearest neighbors. Because several components build on existing Gaussian representation and deformation techniques, a direct comparison is needed to judge both the novelty and the size of the empirical gain.

  1. Limited evidence that the learned heterogeneity is correctly attributed, and limited experimental validation.

The paper claims that one Gaussian framework can represent both conformational and compositional heterogeneity. The reconstruction results are strong, but the experiments do not test whether the learned geometry and amplitude changes correspond to the true structural changes.

Geometry and amplitude are optimized jointly through the reconstruction loss, and nothing in the model explicitly separates them. Amplitude changes could therefore compensate for errors in deformation, and deformation could in turn absorb variation that is actually compositional.

The supplementary Ribosembly analysis shows that amplitude changes are spatially correlated with compositional differences. A spatial correlation, however, does not measure attribution accuracy, nor does it reveal how much the two mechanisms interfere with each other. Synthetic experiments with known motions and known density changes would answer this directly.

On the experimental side, quantitative results cover three synthetic CryoBench datasets, and real-data validation is limited to EMPIAR-10180. The high-frequency improvements there are encouraging, but more experimental datasets are needed to show robustness across molecular systems, heterogeneous states, and imaging conditions.

In short, the paper shows strong reconstruction performance, but offers less support for its broader claims that the representation is physically meaningful and generalizes well.

Questions for authors

Q1. How does HeteroGauss compare with CryoSPIRE and e2gmm?

Could the authors compare against both under the same fixed-pose protocol? CryoSPIRE in particular also uses state-dependent Gaussians coordinated by anchors, so this comparison is the clearest way to show what anisotropic primitives and adaptive topology actually add.

Q2. Can the authors directly test whether geometry and amplitude capture the intended mechanisms?

The Ribosembly analysis shows that amplitude changes align spatially with compositional differences, but it does not show that each mechanism captures only what it is meant to. Using synthetic data with known motions and known density changes, how accurately are each recovered, and how much does one leak into the other?

Q3. Do the gains from adaptive topology come from better allocation or simply from more Gaussians?

Adaptive topology gives large improvements in the ablations, but it likely also changes the final number of Gaussians. How does it compare with a fixed-topology model that has the same final count? This would separate the effect of state-aware allocation from that of added capacity.

Q4. What is the end-to-end benefit of the batched CUDA renderer?

The supplementary material reports higher rendering throughput than per-particle processing, but faster rendering does not necessarily mean proportionally faster training. Could the authors report total training time and peak GPU memory, alongside representative baselines?

Q5. How reliable are the improvements on experimental data?

On EMPIAR-10180, were masking, half-map reconstruction, and post-processing identical for all methods? Do the high-frequency gains over RECOVAR hold across random seeds, and on other experimental datasets?

4 comments
AuthorsOpenReview ↗

Response to reviewer qLTi (Part 1/2)

Thank you for your thoughtful review and for asking how the learned Gaussian changes relate to the underlying heterogeneity and how the gains were validated on experimental data. We respond to the concerns you raised in the following sections.


W1 and Q1

CryoSPIRE. CryoSPIRE is a strong contribution to the field. Unfortunately, its code remains unavailable, and we did not receive a response to our email requesting it, so we left it out of our original experiments. The hardware configurations also differ, which prevents a direct comparison with the results in their paper. To address this question, we reimplemented CryoSPIRE from the paper and ran it on CryoBench. Because the exact experimental settings were unavailable, the numbers reflect our reimplementation under the settings we used. We report them in Appendix B.5 to address your question. The reimplementation scored below both HeteroGauss and the results reported in the CryoSPIRE paper. We attribute the gap to CryoSPIRE's sensitivity to the initial Gaussian parameters. The detailed initialization settings were unavailable. This comparison highlights the advantage of starting from randomly initialized Gaussian parameters in HeteroGauss. The details and visual results can be found in Appendix B.5, or just click here.

CryoBench AU-FSC for the CryoSPIRE reimplementation (revised Table 8).

DatasetCryoSPIRE mean (std)CryoSPIRE medianHeteroGauss mean (std)HeteroGauss median
IgG-1D0.318 (0.002)0.3180.398 (0.001)0.398
IgG-RL0.340 (0.006)0.3400.380 (0.010)0.380
Ribosembly0.347 (0.019)0.3500.430 (0.014)0.432

e2gmm. e2gmm uses reconstruction gradients from the consensus stage as input to its encoder. This makes learning unstable when heterogeneity involves large motions and limits its sensitivity to complex motions. For this reason, we did not use it as a baseline in our main CryoBench experiments. To test this behavior, we also ran e2gmm on IgG-1D and report the results in Appendix B.5. Even with a carefully prepared initialization model, e2gmm did not recover plausible structures across the large rotation in IgG-1D. The details and visual results can be found in Appendix B.5, or just click here.

e2gmm comparison on IgG-1D (revised Table 7).

MethodMean AU-FSC (SD)Median AU-FSCMean time (min/epoch)
HeteroGauss0.398 (0.001)0.3984.46
e2gmm0.2863 (0.0228)0.277830.22

W2 and Q2

Figure 6 in Appendix B.4 of the original submission shows that, with Gaussian centers fixed, amplitude changes alone recover the main regions of ground-truth density change in Ribosembly.

To address your doubts, we added a quantitative analysis of Gaussian motion on IgG-1D. Figure 9 in Appendix B.6 of the revision directly compares the motion of the Gaussian centers with the known Fab rotation in IgG-1D. The mean absolute angular error is 9.92°, in line with the expected motion. You can also click here to view the visual results.These results support the use of amplitude changes to recover compositional differences and of geometric changes to recover domain motion, respectively.

AuthorsOpenReview ↗

Response to reviewer qLTi (Part 2/2)

Q3

We already considered the impact of the number of Gaussians when designing our ablation study. The none-topology model uses the same number of Gaussians as the adaptive model has after topology adaptation. The counts are 12,803 for IgG-1D, 12,034 for IgG-RL, and 32,293 for Ribosembly. With the counts matched, the improvement comes from topology adaptation.


Q4

Table 4 in Appendix B.2 of the original submission shows that, in the At a fixed batch size, reducing rendering time per batch increases the number of particles processed per second by the same factor. The speedup calculated from rendering time therefore equals the speedup calculated from throughput. Revised Table 5 reports both measures.

Furthermore, to further address your concern, we have added a comparison with cryoDRGN, one of the most widely recognized baselines, in Table 6 of the revised appendix, reporting time per epoch and peak GPU memory. Specifically, HeteroGauss trains faster per epoch on both IgG datasets with similar GPU memory use. Moreover, the difference is even larger on EMPIAR-10180. In particular, at a batch size of 128, HeteroGauss trains successfully while cryoDRGN runs out of memory. Even so, HeteroGauss still takes less time per epoch and uses less GPU memory than cryoDRGN at its reduced batch size of 64. The details can be found in Appendix B.3.

Training efficiency (revised Table 6).

DatasetMethodBatch sizeMean time (min/epoch)Peak GPU memory (GiB)
IgG-1DHeteroGauss1284.4615.73
IgG-1DcryoDRGN1285.0814.59
IgG-RLHeteroGauss1284.3014.83
IgG-RLcryoDRGN1285.0914.59
RibosemblyHeteroGauss12836.9838.75
RibosemblycryoDRGN12816.9914.62
EMPIAR-10180HeteroGauss12813.8227.50
EMPIAR-10180cryoDRGN6454.3734.67

Q5

On EMPIAR-10180, HeteroGauss and RECOVAR in Fig. 4D use the same particle partition, soft mask, and frequency grid, while independently reconstructing the half-maps with their respective methods. Similarly, the visualizations in Fig. 4A adopt the same filtering and density display threshold, with the detailed procedure provided in Appendix B.1 of the revision. Furthermore, the two half-map models were trained with different random seeds, so the experimental setup itself already avoids perturbations caused by random seeds. In addition, because preprocessing experimental data, training the half-map models, and evaluating the results are extremely time-consuming, we could not complete further validation during the rebuttal period, but we will supplement results on additional experimental datasets in future work.


Once again, we appreciate the time and effort you spent helping us refine this work, hope our responses have addressed the points you raised, and look forward to your response.

Thank you to the authors for the experimental results and the response; they have addressed my concerns, and I will adjust my score accordingly.

Thank you for your recognition of our work and for raising your score. We are grateful for your thoughtful suggestions, which helped shape the additional experiments and strengthen the paper.

Reviewer_okqQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces HeteroGauss, a state-conditioned anisotropic Gaussian representation for heterogeneous cryo-EM reconstruction, with fixed, pre-estimated particle poses. Specifically, Gaussian positions, scales, orientations, and amplitudes vary to model conformational and compositional variability. Reconstruction quality is evaluated on three CryoBench datasets and the EMPIAR-10180 dataset.

Critical strengths
  1. The spatial anchors and adaptive topology are well-motivated for solving the heterogeneous reconstruction problem.

  2. On IgG-1D and IgG-RL, the proposed method shows clear advancement over previous methods like RECOVAR, and qualitative evaluation also shows that the reconstructions from the proposed method preserve more high-frequency detail.

Critical weaknesses
  1. An essential experimental comparison with CryoSPIRE is missing. Although this paper provides an intuitive comparison with CryoSPIRE in the method part, a comparison with CryoSPIRE is needed in the experiments, as this paper claims to achieve state-of-the-art reconstruction performance.

  2. Batched CUDA rendering is written as a contribution of the method, while there's no efficiency analysis such as timing, throughput, peak memory, or scaling efficiency.

Questions for authors
  1. Why is CryoSPIRE excluded from the comparison in this paper? This is my major concern. It seems that the reconstruction performance of the proposed HeteroGauss is very similar to that of CryoSPIRE according to their paper.

  2. It would be better to have a clear description of the evaluation setting for masked gold-standard FSC in Figure 4. Are particles split before learning the consensus structure, encoder, and decoder, or just at the final inference stage?

I would increase my score if the rebuttal resolve my above concerns.

2 comments
AuthorsOpenReview ↗

Response to reviewer okqQ

Thank you for your careful review and for raising questions about the CryoSPIRE comparison, renderer efficiency, and half-map evaluation. Our responses to your specific questions are below.


W1 and Q1

CryoSPIRE is a strong contribution to the field. However, its code remains unavailable, and we did not receive a response to our email requesting it. We therefore left CryoSPIRE out of the quantitative comparison in the original submission and discussed the differences between the two methods qualitatively. To address your question, we reimplemented CryoSPIRE from the paper and ran it on CryoBench. Because the exact experimental settings were unavailable, we report the results in Appendix B.5. The reimplementation fell short of the published results, and HeteroGauss outperformed it across the board. We attribute this gap to CryoSPIRE's sensitivity to the initial Gaussian parameters and its reliance on a specialized initialization strategy whose exact settings were unavailable to us. HeteroGauss starts from randomly initialized Gaussian parameters, which makes initialization more flexible and robust. The details and visual results can be found in Appendix B.5 or just click here.

CryoBench AU-FSC for the CryoSPIRE reimplementation (revised Table 8).

DatasetCryoSPIRE mean (std)CryoSPIRE medianHeteroGauss mean (std)HeteroGauss median
IgG-1D0.318 (0.002)0.3180.398 (0.001)0.398
IgG-RL0.340 (0.006)0.3400.380 (0.010)0.380
Ribosembly0.347 (0.019)0.3500.430 (0.014)0.432

W2

Thank you for asking for a more detailed efficiency analysis. It should be noted that, Table 5 in the original appendix already reported rendering throughput and peak GPU memory.

However, in order to address your confusion, we added rendering time and measured efficiency across a range of batch sizes. The revised Table 5 is reproduced below. Our method has achieved a comprehensive lead, and the benefits become more pronounced as the batch size increases. The details can be found in Appendix B.3 Table 5 of the revised version.

Renderer efficiency (revised Table 5).

Batch sizeLoop (ms/batch)Batched (ms/batch)Loop (images/s)Batched (images/s)Loop memory (GiB)Batched memory (GiB)Speedup
1624.8911.49642.91,392.10.3520.3502.17×
3246.8420.74683.21,543.20.3870.4042.26×
64102.3538.91625.31,645.00.4650.4902.63×
128218.7675.21585.11,701.80.6190.6662.91×
256508.88149.13503.11,716.70.9221.0233.41×

Q2

We split the particles before learning the consensus structure. Specifically, we use the RECOVAR partition, with 68,599 particle images in each of the two disjoint halves. We then trained HeteroGauss separately on each half for 50 epochs, first fitting its consensus structure and then learning its encoder and decoder. Subsequently, to generate each half-map, we decoded that half's mean latent vector into a 3D density map using its own decoder. Finally, we applied the same RECOVAR mask to both half-maps and computed the FSC. We have added this procedure to Appendix B.1 of the revision.


We thank you again for the time and effort you put into helping us improve our work, hope our responses have addressed your concerns, and look forward to hearing from you.

Dear Reviewer okqQ,

Thank you again for your thoughtful feedback, which has helped us improve both the experiments and the clarity of our paper. We’ve posted a detailed response to your concerns and would be grateful if you could take a look before the discussion period ends in less than two days. We’d be happy to clarify anything further, and if our response resolves your concerns, we would appreciate your reconsideration of the rating.

Reviewer_tw9cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents HeteroGauss, a framework for heterogeneous cryo-EM reconstruction based on a shared representation of state-conditioned anisotropic 3D Gaussian primitives. The method aims to capture both conformational and compositional heterogeneity by allowing the positions, orientations, scales, and amplitudes of Gaussian primitives to vary with the inferred molecular state.

Specifically, a variational encoder infers a latent molecular state from each particle image. As illustrated in Figure 1, the inferred latent representation is combined with anchor features to condition a Gaussian deformation model. Learnable spatial anchors coordinate local motions, while state-dependent amplitude modulation allows Gaussian density contributions to weaken or disappear. The authors further introduce an adaptive topology mechanism that clones, splits, or prunes Gaussians based on reconstruction gradients and state-dependent positional variations. A batched differentiable CUDA renderer is developed to support efficient particle-specific projection and optimization.

Experiments on three CryoBench datasets show improvements over the baselines included in the paper, particularly RECOVAR. On the experimental EMPIAR-10180 dataset, the method also reports improved reconstruction quality and higher spatial resolution.

Critical strengths
  1. Using anisotropic Gaussians to represent the density is a sensible choice for heterogeneous reconstruction. The representation is explicit and quite flexible, and because the Gaussian geometry and amplitude depend on the latent state, the model can in principle capture both conformational and compositional changes. Projecting the Gaussians analytically also fits well with the additive image formation model in cryo-EM.

  2. The adaptive topology mechanism is a useful addition to Gaussian-based reconstruction. Instead of fixing the number and placement of Gaussians in advance, the model can put more of them where the structure needs finer detail. The ablations suggest this component accounts for a large part of the improvement in reconstruction quality.

  3. The results look good on both synthetic and real data. On IgG-1D and IgG-RL the method outperforms the baselines reported in the paper, and the qualitative figures show that flexible regions are better preserved. The EMPIAR-10180 results are also a good sign, especially the gains in high-frequency FSC and local resolution.

Critical weaknesses
  1. The novelty over existing Gaussian-based methods is limited. The overall formulation is close to e2gmm, DynaMight and CryoSPIRE, which have already used Gaussian representations, latent-conditioned deformation, and shared structural elements to model heterogeneity. CryoSPIRE in particular already combines state-conditioned Gaussian geometry, amplitude modulation and anchor-based motion, and it handles both conformational and compositional heterogeneity. It is also missing from the quantitative comparison. It is discussed in both the related work and the method sections but does not appear in Table 1.

  2. The ablations don’t isolate the main contributions. The paper ablates anchor-rigid transport, amplitude modulation, adaptive topology and frequency weighting, but several key design choices are left untested. There is no isotropic vs. anisotropic comparison, even though anisotropy is presented as a main contribution. The adaptive topology ablation also mixes two effects: better spatial allocation and simply having more Gaussians. Comparing against a fixed Gaussian bank with the same final number of primitives would separate the two. In addition, the motion-aware densification should be compared with plain gradient-based densification to show that the state-dependent motion statistics actually help. Without these, it is hard to tell which parts of the design are responsible for the gains.

  3. It is not clear that the learned Gaussian variations reflect the true heterogeneity. The method changes positions, covariances and amplitudes together. That makes the model more flexible, but it also makes the explanation of a given structural change ambiguous: a local density change could come from moving Gaussians, rescaling them, or changing their amplitudes. So I’m not sure the learned amplitude changes correspond to real compositional differences, as opposed to partly compensating for deformation or reconstruction error. Also, keeping Gaussian identities fixed across states does not by itself mean their trajectories correspond to physically meaningful motion. The latent trajectory and density visualizations are helpful, but a quantitative check against the known motions and component occupancies in the benchmarks would make this much more convincing.

Questions for authors

Q1. CryoSPIRE seems to be the closest prior work, so why is it left out of the experiments? Can the authors run a direct comparison using the same datasets, or just use the results reported in their paper?

Q2. Is anisotropy actually needed? An ablation with isotropic vs. anisotropic Gaussians, keeping everything else in the architecture and training fixed, would answer this. How much of the improvement comes from the anisotropic covariance itself, as opposed to the adaptive topology or just more model capacity?

Q3. I’d like to better understand how the particle latent and the anchor features work together. In Figure 1(A), the particle latent is concatenated with anchor features before the deformation decoder. How exactly are the anchor features built and conditioned on the particle latent, and what is the reason for the concatenation? What in the architecture pushes the decoder to actually use the spatially coordinated anchor features, instead of mostly relying on the particle latent? Could the anchor pathway end up being redundant? Ablations with the full model, a version without anchor-feature conditioning, and a version without the direct concatenation of the particle latent would show what each pathway contributes.

Q4. For the gold-standard FSC on EMPIAR-10180, how were the independent half-maps produced? Specifically, were the deformation model, Gaussian parameters and latents optimized separately for each half-dataset? Given how expressive the state-conditioned representation is, what prevents high-frequency information from leaking between halves, or overfitting more generally? Also, the FSC curve never drops to 0.143 within the evaluated frequency range, so how should the reported resolution bound be read?

5 comments
AuthorsOpenReview ↗

Response to reviewer tw9c (Part 1/3)

Thank you for your detailed review and for suggesting ablations and quantitative checks that helped us evaluate the individual components of HeteroGauss. We address your comments point by point below.


W1 and Q1

Novelty. Compared with e2gmm, DynaMight, and CryoSPIRE, HeteroGauss introduces an adaptive topology strategy that improves the spatial allocation of Gaussians and a CUDA rendering kernel that supports batching. All three prior methods rely on specialized tools to initialize their Gaussian parameters. We can train directly from randomly initialized Gaussian parameters, reducing sensitivity to initialization and making training more flexible and efficient.

CryoSPIRE. CryoSPIRE is a strong contribution to the field. However, its code remains unavailable, and we did not receive a response to our email requesting it, so we left it out of our original experiments. Moreover, the hardware configurations also differ, which prevents a direct comparison with the results in their paper. Nevertheless, to address this question, we reimplemented CryoSPIRE from the paper and ran it on CryoBench. Because the exact experimental settings were unavailable, the resulting numbers reflect our reimplementation under the settings we used. We therefore report them in Appendix B.5 for your reference**. **In these experiments, the reimplementation scored below both HeteroGauss and the results reported in the CryoSPIRE paper. We attribute this gap to CryoSPIRE's sensitivity to the initial Gaussian parameters, and the detailed initialization settings were also unavailable. Ultimately, this comparison highlights the advantage of starting from randomly initialized Gaussian parameters in HeteroGauss. The details and visual results can be found in Appendix B.5 or just click here.

CryoBench AU-FSC for the CryoSPIRE reimplementation (revised Table 8).

DatasetCryoSPIRE mean (std)CryoSPIRE medianHeteroGauss mean (std)HeteroGauss median
IgG-1D0.318 (0.002)0.3180.398 (0.001)0.398
IgG-RL0.340 (0.006)0.3400.380 (0.010)0.380
Ribosembly0.347 (0.019)0.3500.430 (0.014)0.432

e2gmm. e2gmm uses reconstruction gradients from the consensus stage as input to its encoder. This makes learning unstable when heterogeneity involves large motions and limits its sensitivity to complex motions. For this reason, we did not use it as a baseline in our main CryoBench experiments. To test this behavior, we also ran e2gmm on IgG-1D and report the results in Appendix B.5. Even with a carefully prepared initialization model, e2gmm did not recover plausible structures across the large rotation in IgG-1D. The details and visual results can be found in Appendix B.5 or just click here.

e2gmm comparison on IgG-1D (revised Table 7).

MethodMean AU-FSC (SD)Median AU-FSCMean time (min/epoch)
HeteroGauss0.398 (0.001)0.3984.46
e2gmm0.2863 (0.0228)0.277830.22
AuthorsOpenReview ↗

Response to reviewer tw9c (Part 2/3)

W2 and Q2

Thank you for asking which components drive the gains and how we controlled the ablations. We added two comparisons and clarified how we matched the Gaussian counts.

  • Is anisotropy necessary? We compared anisotropic and isotropic Gaussians on all three CryoBench datasets. With isotropic Gaussians, AU-FSC decreases on all three datasets (revised Table 2). The reconstructions are also smoother and lose high-frequency detail compared with those from the anisotropic model (Fig. 5).
  • Does the gain come from topology or from greater model capacity? We already considered the impact of the number of Gaussians when designing our ablation study. The none-topology model uses the same number of Gaussians as the adaptive model has after topology adaptation. The counts are 12,803 for IgG-1D, 12,034 for IgG-RL, and 32,293 for Ribosembly. With the counts matched, the improvement comes from topology adaptation.
  • Motion-aware vs. gradient-only topology. To test the contribution of motion information, we added a gradient-only ablation. Using gradients alone lowers AU-FSC on all three datasets (revised Table 2). In Fig. 5B, the gradient-only reconstruction also shows a detached density fragment in the lower left. These results support using motion information to guide topology adaptation.

You can also click here to view the visual results.

Relevant component ablations (revised Table 2). The table reports mean AU-FSC with standard deviation and median across conformations.

VariantIgG-1D mean (std)MedianIgG-RL mean (std)MedianRibosembly mean (std)Median
Full HeteroGauss0.398 (0.001)0.3980.380 (0.010)0.3800.430 (0.014)0.432
Isotropic Gaussians0.389 (0.002)0.3890.368 (0.009)0.3680.410 (0.012)0.414
Gradient-only densification0.390 (0.002)0.3900.371 (0.001)0.3710.422 (0.014)0.424

W3

The Gaussians change with the latent embeddings. Figure 3 in the original submission showed that these embeddings track the ground-truth conformational changes, so the resulting Gaussian variations reflect the underlying heterogeneity. Figure 6 in Appendix B.4 also showed that the learned amplitude changes correspond to ground-truth compositional differences.

Following your suggestion, we added a quantitative analysis of Gaussian motion on IgG-1D. We tracked the same group of Fab Gaussians across conformations and compared each fitted rotation angle with its ground-truth value. The fitted angles follow the Fab through a full rotation, with a mean absolute angular error of 9.92° (Appendix B.6, Fig. 9). This shows that the tracked Gaussians recover the overall motion of the Fab domain. The details can be found in Appendix B.6 of the revised version.You can also click here to view the visual results.

AuthorsOpenReview ↗

Response to reviewer tw9c (Part 3/3)

Q3

The “concat” label in the original Fig. 1(A) was meant to show how the particle latent and anchor features work together to control the Gaussian parameters. The label made it look as though the two were directly concatenated, so we revised the figure to clarify the data flow. As described in Appendix A.2 of the original submission, the network uses the particle latent to generate a feature for each anchor. Each Gaussian uses its position in the canonical structure to find its four nearest anchors and computes a distance-weighted average of their features. The decoder concatenates this local feature with the positional encoding of the Gaussian’s own center. This lets neighboring Gaussians share state information from nearby anchors while retaining information about their individual positions.

Based on above,the interpolated anchor feature is the only source of particle-state information for the per-Gaussian update heads. The full model already corresponds to the suggested variant without direct latent concatenation, so those two settings are identical. Removing the anchor features would leave the update heads with only positional encodings, so their predictions would no longer depend on the particle state. The particle latent and the anchor features are successive steps in the same conditioning pathway. The proposed three-way ablation therefore cannot isolate the contributions of two separate pathways.


Q4

We have added the half-map generation procedure for Fig. 4D to Appendix B.1. Specifically, following the RECOVAR partition, we split the particles into two disjoint halves of 68,599 images each before fitting the consensus structure. We then trained HeteroGauss from scratch on each half, independently learning the consensus structure, Gaussian parameters, encoder, and deformation decoder. Subsequently, each encoder inferred latents for its own half. Next, we generated each half-map by decoding the mean latent of that half with its own decoder, applied the same RECOVAR soft mask to both maps, and computed the FSC. Thus, because each model learns its parameters from its own half, the two models have no shared weights through which high-frequency noise could pass.

Regarding the resolution plot, the horizontal axis in Fig. 4D represents spatial frequency. Accordingly, the farther to the right the FSC curve crosses 0.143, the higher the spatial frequencies at which the two half-maps agree and the finer the corresponding structural detail. In our results, HeteroGauss remains above 0.143 at the highest frequency we evaluated, so the agreement extends to the limit of the evaluated frequency range and the threshold crossing lies beyond that range. Specifically, the highest sampled frequency corresponds to 4.31 Å, we therefore report ≤4.31 Å.


We thank you once more for the time and care you devoted to helping us strengthen our work, hope our replies have answered your questions, and look forward to your feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Reply to the Area Chair and Reviewers

We thank the Area Chair for considering our submission and the reviewers for their time and detailed feedback. Several questions came up across the reviews, especially about the choice of baselines, the evidence for our design choices, and the evaluation protocol. We have marked the revisions in blue in the manuscript and included Anonymous GitHub links in our responses so reviewers can easily view the results while preserving double-blind review. Below, we summarize these shared concerns and how we addressed them in the revision.

  • CryoSPIRE comparison. All three reviewers raised the missing comparison with CryoSPIRE as their first concern. Because CryoSPIRE’s code is unavailable, we discussed the method qualitatively in the original submission and omitted it from the quantitative comparison. To address this concern directly, we reimplemented CryoSPIRE from the paper and tested it on CryoBench. HeteroGauss outperformed this reimplementation on all three CryoBench datasets reported in Table 8 (Appendix B.5).
  • e2gmm comparison. We left e2gmm out of the main CryoBench comparison because its encoder relies on gradients computed at the consensus structure, which limits its ability to capture large conformational motions. To test this limitation, we ran e2gmm on IgG-1D. Despite careful initialization, it failed to recover the large Fab rotation (Appendix B.5).
  • Component ablations. We added isotropic and gradient-only variants. Both reduced AU-FSC across the three CryoBench datasets. We also clarified that the fixed-topology variant matches the final Gaussian count of the adaptive model on each dataset (Table 2).
  • Efficiency. To address the reviewers’ concerns about efficiency, we extended our original analysis by benchmarking the renderer across batch sizes and comparing HeteroGauss and cryoDRGN on training time per epoch and peak GPU memory (Tables 5 and 6).
  • Gold-standard FSC on EMPIAR-10180. In Appendix B.1, we clarified how we split the particles, trained a separate model on each half, and computed FSC from the resulting half-maps.

We address the remaining questions and provide additional details in our individual replies to each reviewer.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.