acceptodds
Under review as a conference paper at ICLR 2027

CryoSplat2: Gaussian Splatting for Cryo-EM Ab Initio Reconstruction

Abstract

Cryogenic electron microscopy (cryo-EM) is a widely used structural biology technique for determining the 3D structures of biomolecules from noisy 2D particle images. In single-particle cryo-EM, *ab initio* reconstruction jointly estimates 3D density maps and unknown particle poses without an initial 3D reference, making the problem severely ill-posed due to the strong coupling between pose estimation and structure reconstruction. Gaussian representations provide an explicit, continuous, and differentiable model of molecular density, with adaptive degrees of freedom and efficient projection, but most existing Gaussian-based cryo-EM methods assume known particle poses. We introduce cryoSplat2, a Gaussian-based framework for particle-level *ab initio* reconstruction that jointly optimizes particle poses and anisotropic Gaussian representations from random initialization. In particular, cryoSplat2 combines frequency marching for coordinated coarse-to-fine pose and structure refinement, adaptive Gaussian densification for progressively increasing representation capacity, and an optimized differentiable splatting operator for efficient pose inference. Experiments on multiple real cryo-EM datasets show that cryoSplat2 robustly reconstructs 3D structures without an initial 3D reference or externally provided particle poses, including challenging cases where cryoSPARC or cryoDRGN-AI fails to converge to the correct structure. The code will be released upon paper publication.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Cryo-EM, 3D Reconstruction, Gaussian Mixture Model, Inverse Problems

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_PVx7OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes cryoSplat2, a Gaussian-based framework for particle-level ab initio reconstruction. Specifically, it jointly optimizes particle poses and anisotropic Gaussian representations from random initialization. Experiments include four real cryo-EM datasets, evaluating ab initio reconstruction quality, downstream refinement, and computational efficiency.

Critical strengths
  1. This work successfully demonstrates that a Gaussian-based method can also achieve ab initio cryo-EM reconstruction, which can bring significant efficiency advances over cryoDRGN.

  2. Experiments show that the proposed method can produce reliable reconstruction across diverse structures, including challenging cases.

Critical weaknesses
  1. The method's key designs such as the scheduler and Gaussian densification are not separately ablated to show their individual effects.

  2. Robustness and reliability need more justification. In the experiments, this paper reports some failures of prior works and some successful runs. It would be more clear to report the number of runs and success rates of all methods to have a clear robustness and reliability evaluation.

Questions for authors

Is it possible to have controlled comparisons of adaptive versus predetermined scheduling and adaptive versus fixed Gaussian capacity?

Some related work on Gaussian-based cryo-EM methods is missing, such as GEM [1].

[1] Gem: 3d gaussian splatting for efficient and accurate cryo-em reconstruction, Qu, et al., 2025

5 comments

We thank the reviewer for the constructive comments. We conducted the requested additional experiments, incorporated the new results into the revised manuscript, and added the missing related work. We address each point below.

AuthorsOpenReview ↗

W2: Success rate

We have evaluated cryoSplat2, cryoSPARC, and cryoDRGN-AI using five random seeds across all four datasets, with the results added to Appendix C (Figs. 10 and 11). cryoSplat2 succeeds in all 20 runs (Fig. 10). Both baselines succeed in all five runs on EMPIAR-10028 and EMPIAR-11910. However, cryoSPARC fails in all five runs on EMPIAR-10180 (0/5) and succeeds in only two out of five runs on EMPIAR-11762 (2/5), while cryoDRGN-AI fails in all five runs on EMPIAR-11762 (0/5) (Fig. 11). These results demonstrate the greater robustness of cryoSplat2 to random initialization across the evaluated datasets.

AuthorsOpenReview ↗

W1, Q1: Adaptive versus predetermined scheduling

Thank you for suggesting this comparison. We added a controlled ablation that uses a predetermined schedule to emulate an ad hoc fixed frequency-marching workflow. Specifically, we record the stage transitions from a successful EMPIAR-10028 reconstruction using our adaptive frequency-marching scheduler, freeze them, and apply the resulting schedule to EMPIAR-11762, EMPIAR-11910, and EMPIAR-10180 without dataset-specific retuning. The fixed schedule uses the same joint-iteration counts for increasing the HEALPix order and reconstruction bandwidth, switching from global to local pose search, and inserting fixed-pose epochs. Because the datasets have different image sizes and pixel sizes, the Fourier crop side lengths are accordingly converted to preserve the corresponding nominal physical resolutions of the source schedule.

The fixed schedule does not generalize reliably to the other datasets: EMPIAR-11910 exhibits pronounced structural distortion, EMPIAR-11762 misses part of the structure, and EMPIAR-10180 also shows clear degradation. In contrast, our adaptive frequency-marching scheduler consistently produces complete and well-formed reconstructions across all four datasets.

These results show that a fixed coarse-to-fine schedule tuned for one dataset is insufficient for robust cross-dataset reconstruction, further demonstrating the effectiveness and robustness of our adaptive scheduler. We added this experiment and the corresponding fixed schedule to Appendix C under “Necessity of adaptive frequency marching.”

Reviewer_zxbJOpenReview ↗
Rating
3
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a method for ab initio Cryo-EM reconstruction based on Gaussian splatting. The work follows the recent CryoSplat method, but builds on it by lifting the assumption of known poses, and hence enabling ab initio reconstruction. To estimate poses, the approach introduces a frequency marching optimization along with a coarse-to-fine pose search, where both structure and poses are optimized jointly. Additional improvements to the Gaussian splatting model are introduced to significantly speed up the rendering and optimization procedure compared to previous work. The method is evaluated on multiple particle datasets and compared to other methods, including CryoSPARC and cryoDRGN-AI. Overall, it performs similarly in terms of reconstructed resolution for most datasets, but it offers substantial computational speedups compared to cryoDRGN-AI. Compared to CryoSPARC it appears to offer some advantages in terms of robustness of the optimization. Additional evaluations included in the Appendix provide some exploration of extensions to the heterogeneous case via classification.

Critical strengths

Overall, this is a well-executed paper and the evaluation and descriptions are quite thorough, with many additional details provided in the appendices.

  • The method is evaluated on multiple real datasets and compared to relevant related techniques, and it appears to provide some advantages over previous work.
  • Further, the speedups relative to cryoDRGN-AI are compelling, and it may offer some advantages compared to cryoSPARC in terms of the robustness of the optimization. For example, the authors show a case where the cryoSPARC optimization appears to fail — though it's not entirely clear how common this failure mode is — as well as some results that argue for the benefit of the GMM representation vs. a conventional voxel grid.
  • The additional extensions to classification are interesting and show the method could extend to enable classification and consensus refinement.
Critical weaknesses

Although the method offers some improvement in results compared to baselines, I didn't come away from the paper seeing very significant new technical innovations. I think this is because the method largely relies on applying existing techniques to get the GMM representation working for ab initio reconstruction. To some extent each of the technical contributions has been demonstrated in previous work, though perhaps not exactly in this particular setting.

  • Adaptive densification is common in the Gaussian splatting literature, and here the authors adapt it to the cryo-EM case. The particular implementation here is more complex than what was shown in cryoSplat, but the paper doesn't thoroughly evaluate how much the choices in the densification algorithm affect model performance.
  • Frequency marching and hierarchical pose search are well-established techniques in the cryo-EM literature, and the authors cited relevant papers that already incorporate these techniques.
  • The optimization of the Gaussian splatting operations largely conform to what has been shown in previous work (e.g., gsplat, Speedy-Splat, and Taming 3DGS), as the authors discussion in Section 3.5.

It's nice to see these techniques brought together, and they work well for the application; however, I don't see any specific contribution as contributing something that is substantially new or strikingly different than previous work.

Questions for authors

I have a few additional comments/questions.

  • It would be helpful to put some details of the pruning into the main paper (such as an equation that gives the gist of the main pruning criterion).
  • How sensitive is the method to the initialization of Gaussians, and how consistent is run-to-run performance?
  • How much does performance depend on the number of Gaussian primitives used? Is there any reason not to use many more Gaussians vs fewer Gaussians — maybe performance levels out (or worsens) once a certain number of Gaussians is used?
  • Could the authors say more about why the Gaussian representation is more robust to smaller batch sizes than the voxel representation? Are there specific properties of the Gaussian representation that contribute here?
10 comments

Thank you for the thoughtful comments and questions. We address each point in order below.

AuthorsOpenReview ↗

W1: Adaptive Gaussian densification

We agree that densification itself is widely used in other Gaussian splatting works, and we do not claim novelty in the general idea of increasing Gaussian capacity. The key difference in cryoSplat2 is that densification is coupled directly to the current reconstruction bandwidth: Gaussian primitives that are too coarse for the nominal spatial resolution are split using a moment-preserving construction, so the representation capacity is progressively released as finer structure becomes resolvable.

To better evaluate the role of this design, we added two complementary ablations. In the first experiment, we vary the initial Gaussian count while keeping the adaptive densification enabled. Except for extremely under-capacitated initializations with only one or two Gaussians, most initial counts converge to visually similar structures, showing that adaptive densification substantially reduces sensitivity to the initial representation capacity (Fig. 12). In the second experiment, we disable densification and use fixed Gaussian counts ranging from 512 to 32,768. We find that simply allocating more Gaussians from the beginning does not necessarily improve reconstruction quality. For example, a configuration initialized with 32,768 Gaussians recovers substantially less structural detail than adaptive densification, despite using more primitives than the 20,624 Gaussians in the final densified model (Fig. 13). These experiments show that the benefit of densification is not merely increasing the final Gaussian count, but releasing additional degrees of freedom at appropriate stages of coarse-to-fine optimization. We have added these results to Appendix D.

AuthorsOpenReview ↗

W2: Frequency marching and hierarchical pose search

We agree that frequency marching and hierarchical pose search are established techniques in cryo-EM. We do not claim novelty in either technique individually. Rather, our contribution lies in developing a unified optimization framework that integrates these techniques with the Gaussian representation for ab initio reconstruction. Specifically, cryoSplat2 jointly coordinates reconstruction bandwidth, pose-search resolution, and Gaussian representation capacity through a unified adaptive scheduler, enabling joint pose--structure optimization from random initialization. This coordination is important because increasing spatial bandwidth requires both sufficiently accurate pose estimates and adequate Gaussian capacity. We have revised the paper to clarify this contribution and expanded the scheduler implementation details in Appendices C and F.

Reviewer_JA4QOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper builds on prior work using splatting procedures to represent cryo-EM volumes for ab-initio reconstruction. They introduce a pose marginalization technique for handling unknown pose, in contrast to the common approach of using a maximum a posteriori estimate. They then test the benchmark against various approaches, and perform ablation tests on the Gaussian splatting approach.

Critical strengths

The paper is well written, clearly presented. I found the experiments convincing, and well thought out. Frankly, I have been pretty skeptical of splatting strategies in cryo-EM before this but this paper convinced me that I should give them another look.

Critical weaknesses
  • I am a little cautious about some of the baselines – I am surprised that cryosparc performs so poorly on EMPIAR-10028. Do the authors have an explanations on why this dataset failed?
  • In addition to seeing an ablation of the Gaussian representation, I’d like to see an ablation where the max likelihood pose was used instead of pose marginalization. I think the use of a pose marginalization is an underemphasized contribution of this paper and, a little more experiments would merit highlighting it as an additional contribution in this paper.
Questions for authors
  • For the comparison with voxel-based strategies, is it clear that the correct frequency marching parameters are appropriate in both cases? For that matter, how is the frequency marching applied to voxel representation? I was unable to find these details in the paper. In the interests of full disclosure, these questions are the primary one keeping me from a weak to clear acceptance: with this point clarified I am happy to reconsider my rating.
7 comments

Thank you for the thoughtful comments and for highlighting aspects of the paper that could be clarified further. We have conducted the requested pose-marginalization ablation and revised the manuscript accordingly. We address each point below.

AuthorsOpenReview ↗

W1: Performance of cryoSPARC

We would first like to clarify that cryoSPARC does not perform poorly on EMPIAR-10028 in our experiments. As shown in Fig. 2, all three methods successfully reconstruct this dataset and produce comparable downstream refinement results. We believe the reviewer may instead be referring to EMPIAR-10180.

The difficulty of EMPIAR-10180 has also been observed in prior work. CryoDRGN2 reported that successful reconstruction on this dataset required recentering the particles using published poses. As noted in Appendix H, the deposited EMPIAR-10180 particles contain substantial in-plane translations. We suspect that these large translations, together with substantial structural heterogeneity, make joint pose and structure optimization particularly challenging. Because our goal is fully ab initio reconstruction, we use the deposited particles directly and do not recenter them using published poses. We have clarified this point in Sec. 4.2 and Appendix H (Dataset Details).

AuthorsOpenReview ↗

W2: Pose marginalization vs. maximum-posterior pose

We agree that pose marginalization is an important aspect of the reconstruction formulation and thank the reviewer for suggesting this ablation. We conducted an additional experiment in which the pose search and posterior computation remain unchanged, but only the maximum-posterior pose for each particle is retained during backpropagation.

Across EMPIAR-10028, EMPIAR-11910, EMPIAR-11762, and EMPIAR-10180, we do not observe a substantial qualitative difference: both full pose marginalization and maximum-posterior updates converge to visually similar structures. We have added these results to Appendix C (Fig. 8). We therefore do not claim that marginalization is necessary for successful Gaussian-based reconstruction on these benchmarks. Interestingly, when using maximum-posterior updates, the voxel-grid representation produces noisier reconstructions, with noticeably degraded results on EMPIAR-11762 and EMPIAR-10180 (Fig. 8). This is consistent with the greater sensitivity of voxel optimization to noisy gradients observed in Sec. 4.3, whereas the Gaussian representation appears considerably more robust to hard, imperfect pose assignments. Nevertheless, we expect it to become more important in more ambiguous regimes where the pose posterior is broad or multimodal and hard assignment may prematurely commit particles to incorrect pose modes. Moreover, pose marginalization in cryo-EM 3D reconstruction has been widely adopted in conventional methods, including RELION and cryoSPARC. Since our implementation primarily involves engineering adaptations, we do not consider pose marginalization itself as a major contribution.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.