acceptodds
Under review as a conference paper at ICLR 2027

CryoSD: Score Distillation for Heterogeneous ab initio Reconstruction in Cryo-EM

Abstract

Heterogeneous *ab initio* reconstruction in single-particle cryo-electron microscopy (cryo-EM) aims to recover an ensemble of 3D molecular structures from extremely noisy 2D projection images. The task is further complicated by unknown particle orientations and contrast transfer function (CTF) distortions. Existing approaches typically combine variational autoencoders with explicit pose estimation, which becomes unreliable in the presence of severe noise, pronounced conformational variability, and structural symmetries. We introduce cryoSD, a framework for heterogeneous cryo-EM reconstruction that combines ideas from score-based diffusion models, Noisier2Noise and score distillation sampling (SDS). CryoSD learns a score-based diffusion model over particle images and leverages the learned image distribution for unsupervised denoising, CTF correction and SDS-based 2D-to-3D reconstruction, eliminating the need for explicit pose estimation. We evaluate cryoSD on four challenging datasets exhibiting continuous and compositional heterogeneity, including IgG-RL and IgG-1D from CryoBench and data from EMPIAR-10295. We compare cryoSD with the state of the art in heterogeneous *ab initio* reconstruction. Our results demonstrate the feasibility of SDS-based heterogeneous cryo-EM reconstruction and suggest cryoSD as a powerful alternative to existing approaches in heterogeneous cryo-EM *ab initio* reconstruction. In particular, cryoSD shows superior performance on sparse heterogeneous data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
score-based diffusion models, score distillation sampling, inverse problems, cryo-electron microscopy, heterogeneous 3D reconstruction, self-supervised denoising, computational imaging

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_NZ7wOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CryoSD, a score-based framework for heterogeneous ab initio cryo-EM reconstruction. It first learns a diffusion model from raw particle images and uses it for joint denoising and approximate CTF correction. A second diffusion model is trained on the restored images. For each selected particle, a 3D volume is then optimized by matching a fixed-view projection to that particle as an anchor, while encouraging projections from other directions to agree with the learned image distribution. Experiments include synthetic conformational heterogeneity datasets and real clathrin-cage data from EMPIAR-10295, with comparisons against cryoDRGN-AI and RELION.

Critical strengths
  1. A novel and intuitive reconstruction framework. Combining a particle-specific conformational anchor with a diffusion prior over other projection directions provides a substantially different approach to heterogeneous reconstruction without explicit rotation estimation. The method offers an interesting alternative to conventional pose-search-based pipelines. The paper is also clearly written and easy to follow, with a coherent introduction that provides useful tutorial-style explanations.

  2. Meaningful evidence from real experimental data. The EMPIAR-10295 experiment provides the strongest evidence for the method’s effectiveness. The recovered volumes show recognizable cage architectures, supporting the practical feasibility of the framework beyond synthetic data. These results make the approach worth further investigation despite the remaining limitations.

Critical weaknesses
  1. A key ab initio baseline is missing. cryoSPARC is a widely used and particularly strong baseline for reference-free initial model generation and multi-class ab initio reconstruction. Comparison with RELION does not replace this evaluation on EMPIAR-10295. This omission is especially important because cryoSPARC provides a high-symmetry preset addressing failure modes directly relevant to cage-like structures without requiring a specified symmetry group. Consequently, the current comparisons do not sufficiently establish an advantage over strong conventional ab initio workflows.

  2. Generality and downstream utility remain insufficiently validated. EMPIAR-10295 contains unusually large cage-like assemblies whose characteristic architectures can be distinguished at relatively coarse resolution. This reduces the dependence on accurately recovering high-frequency information, where CTF oscillations are denser and signal attenuation makes recovery more difficult. This is nevertheless a challenging dataset: its symmetry and pseudosymmetry create substantial ambiguity for ab initio pose search and can lead to collapsed viewing-direction assignments and distorted reconstructions. However, this combination of large particle size and pronounced symmetry represents a specialized setting. Evaluation on EMPIAR-10180 for conformational variability and EMPIAR-10076 for compositional variability in a more compact, asymmetric assembly would provide more convincing evidence of broader applicability.

    Limited resolution is not inherently a weakness for an ab initio method; however, whether the recovered volumes are sufficiently accurate to support subsequent classification and high-resolution refinement has not been demonstrated. Using these volumes as initial references in cryoSPARC or RELION, and successfully recovering the corresponding heterogeneous states at higher resolution, would provide strong evidence of their structural accuracy and practical utility.

Questions for authors
  1. Can the authors compare against cryoSPARC multi-class ab initio reconstruction on EMPIAR-10295, including its high-symmetry preset, using the same particle set without providing known structures or symmetry groups? This is particularly relevant because the paper attributes its advantage partly to difficulties faced by pose-search-based methods on symmetric or pseudosymmetric structures.

  2. Can the authors provide real-data evaluation beyond large cage-like assemblies whose architectures are distinguishable at coarse resolution? EMPIAR-10180 and EMPIAR-10076 would be informative benchmarks for conformational and compositional variability, respectively. These experiments would help establish whether the method generalizes beyond the specialized combination of large particle size and pronounced symmetry in EMPIAR-10295.

  3. Can the recovered volumes serve as initial references for downstream classification and refinement in cryoSPARC or RELION? High resolution is not essential at the ab initio stage, but successful separation and high-resolution refinement of the corresponding heterogeneous states would provide strong evidence that the initial reconstructions are sufficiently accurate and practically useful.

My current recommendation is clear acceptance, based on the novelty of the framework and the evidence of its effectiveness on real data. However, the missing cryoSPARC comparison remains a substantive concern, particularly given the paper’s claims about advantages over pose-search-based methods. If this concern is not adequately addressed through additional experiments, a convincing explanation, or appropriately narrowed claims, I may revise my recommendation to weak acceptance. The other suggested experiments would strengthen the paper, but I do not consider them prerequisites for acceptance.

2 comments

We thank the reviewer for carefully reading our paper and for preparing this thoughtful review. We regard the reviewer's critical comments and questions as highly valuable guidance for our continued work on this project.

We agree with the substantive concern regarding the missing comparison with CryoSPARC. We have revised the Results section to include CryoSPARC (see Figure 5 in the revised manuscript). This concern was clearly justified: CryoSPARC’s multi-class ab initio reconstruction outperformed RELION and CryoDRGN-AI. CryoSPARC recovered the mini coat architecture and most of the D6 barrel, but did not recover the remaining three architectures. Further details are provided in Appendix E.2 of the revised manuscript.

We sincerely thank the reviewer for identifying this important missing part and for helping us improve the paper.

We also include preliminary qualitative results on the experimental EMPIAR-10180 spliceosome dataset in Appendix E.3 of the revised manuscript.

Thank you for your efforts in addressing my concerns. After carefully reading the revised version and considering other reviewers' comments, I have decided to maintain my current rating.

The results on EMPIAR-10180 largely address my concerns. Although they are somewhat low-resolution, I believe the reconstructed models can be used as initial models for downstream refinement. That is why I highly recommend that the authors compare downstream refinement or 3D classification results using the initial models generated by the proposed method and other baselines. This would provide direct evidence that the initial models are good enough for downstream applications, which I believe could also address other reviewers' concerns.

In my experience, limited resolution or potential generative artifacts are not necessarily fatal issues for ab initio reconstruction results. In downstream RELION-like Bayesian EM refinement, the density is re-estimated directly from experimental particle images rather than inherited from the generated initial models. Therefore, generative artifacts are not directly preserved in the reconstructed density. Although inaccurate initial models may lead to incorrect convergence, I would expect the refinement to either converge toward the correct structures or produce obviously incorrect reconstructions that are unlikely to be mistaken for valid protein structures.

In cryoSPARC, downstream refinement typically takes only a couple of minutes. That means the authors still have a chance to complete these experiments by the end of the rebuttal phase.

Reviewer_KoGAOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The authors present an innovative approach for resolving single-particle cryo-EM reconstructions via a multi-stage approach that takes 2D projection images and generates an ensemble of 3D structures. The method relies first on a 2D U-Net that is trained with denoising score matching. The model is then used to denoise and approximately CTF-correct the images. They then train a score model on the denoised data and use it together with the denoised 2D images to generate 3D reconstructions. This is based on an unconditional 2D score approach that generates a separate 3D volume for each projection image.

Critical strengths

The approach presented in this manuscript is innovative and exposes an orthogonal direction rarely explored for cryoEM reconstructions.

Critical weaknesses

The main contextual weakness is that the authors ignore the fact that most cryoEM SPA datasets contain significant portions of false positives. Both experimental and synthetic datasets are clean datasets that are pruned to exclusively clean particles.

The selected experimental dataset has unusually higher SNR and/or unusually strong low-resolution signal than most cryoEM dataset. Moreover it has substantial structural symmetry that will introduce significant redundancy in the projection images, which is highly relevant for evaluating this method.

The main technical weakness is the unconditional 2D score model used to generate 3D reconstructions from single projection images. Generated 3D volumes are only anchored in a single projection and severely underconstrained. The other views are only constrained by the learnt distribution. If the data has representative noise, structurally complex and heterogeneous as a realistic cryoEM dataset this lack of robustness might impact the reliability of the method.

Questions for authors
  1. Can the authors evaluate the consistency of other projections than the anchor of the reconstructed 3D volumes with existing projection images? They should at least report an FRC.
  2. Can the authors do comparisons to fixed pose SOTA methods?
  3. Can the authors demonstrate the approach on at least one typical dataset with heterogeneity? They can even use an easy ribosome high SNR dataset.
  4. The authors need to demonstrate robustness to “garbage particles”. Can they benchmark the approach with various portions of bad particles?
6 comments
AuthorsOpenReview ↗

Response to Reviewer KoGA

We thank the reviewer for recognizing that our manuscript is innovative and explores an orthogonal direction for cryo-EM reconstruction. We are confident that we can address the concerns raised, as detailed below.

False positives and preprocessing (Q4). In a typical cryo-EM pipeline, particles picked from micrographs undergo 2D classification and averaging, after which the user can remove a large proportion of false positive picks. Like other 3D ab initio reconstruction algorithms, our pipeline starts after this preprocessing step. The input stacks therefore do not contain the substantial proportion of false positives present immediately after particle picking.

A small number of bad particles can remain after preprocessing, as is also the case for EMPIAR-10295, on which our method showed superior performance. As expected, bad particles do not yield reasonable volumes when used as anchors, but they do not prevent reconstruction from other anchors.

Our two-stage approach also offers an advantage that we have not yet highlighted in the paper: after denoising and CTF correction, users can remove false positives or preselect conformations of interest before starting 2D-to-3D distillation. Such an intermediate filtering step is not available within CryoDRGN-AI.

Dataset choice and scope of the contribution (Q3). We do not claim general superiority over CryoDRGN-AI. We propose an orthogonal approach to heterogeneous ab initio reconstruction and demonstrate its potential advantages. Establishing general superiority over the state of the art would require much more evidence; that is not our claim. Pose-search-based approaches such as in RELION, CryoSPARC and CryoDRGN-AI build on more than a decade of substantial research effort. We aim to explore a new direction and share it with the community. New ideas need time to mature and should not be expected to immediately outperform methods across all datasets which are developed and optimized over more than a decade in every setting.

EMPIAR-10295 also has a small particle count, pronounced heterogeneity, and cage architectures with different symmetries or challenging pseudo symmetries. The presence of symmetry does not automatically mean that every method can exploit the resulting redundancy. This dataset is too challenging for methods such as CryoDRGN-AI or RELION. Difficulty for estabilshed methods does not make a dataset untypical. We are interested in datasets that challenge existing approaches, rather than only those for which the community already has effective methods. We also demonstrate that CryoSD works at low SNR on IgG-1D and with realistic, strongly colored noise on capsidHA.

Single-image anchors and 3D constraints. CryoDRGN-AI also produces a 3D volume for each individual noisy projection. That projection determines the conformation encoded by its latent vector. The remaining 3D information comes from the decoder, which learns from the other particle images. In CryoSD, the clean anchor image similarly fixes the conformation, while the remaining 3D information comes from the score model trained on the clean projection images. The score model can be interpreted as a convenient transformed representation of our dataset. It is trained exclusively on the single particle stack, not across different datasets.

We therefore firmly disagree with the characterization of our reconstructions as severely underconstrained. Per-image FSC measures the correlation between each reconstructed volume and the exact ground truth conformation depicted in its input projection. The reported per-image FSC results thus directly counter this concern on the tested datasets, which include low SNR, realistic cryo-EM colored noise, complex structures, and strong heterogeneity.

Q1: Consistency beyond the anchor view. Per-image FSC evaluates consistency directly against the corresponding ground-truth conformation in 3D. It therefore provides a more comprehensive assessment than FRC against other existing projection images.

Q2: Comparisons with fixed-pose methods. The fixed-pose setting in CryoBench assumes that all particle orientations and translations are known a priori. This artificial tomography-like setting is fundamentally easier than, and different from, the ab initio problem addressed here.

I agree that per-image FSC against the ground-truth 3D conformation is a stronger evaluation than FRC against individual projections. Figures 3 and 4 demonstrate that the method can recover the correct conformations on the synthetic datasets. However, Figure 5 evaluates selected reconstructions against published EMDB maps rather than per-particle ground truth.

My main technical concern remains that the unconditional 2D score model marginalizes over orientations and conformations and therefore does not explicitly enforce cross-view conformational consistency. The authors themselves acknowledge cases where the reconstructed volume is plausible but does not correspond to the conformation represented by the anchor.

While the FSC results address my concern regarding the evaluation of cross-view consistency on synthetic data, the underlying methodological concern remains unresolved. My other concerns also remain largely unaddressed, particularly robustness to imperfect restoration, asymmetric structures, and realistic experimental particle distributions. The authors have also not provided the requested comparison against fixed-pose heterogeneity methods using conventionally estimated poses, or a quantitative evaluation of robustness to bad particles.

We thank the reviewer for the follow-up and for acknowledging that per-image FSC addresses cross-view consistency on the synthetic datasets.

We are therefore surprised that the reviewer nevertheless maintains a technical concern regarding cross-view conformational consistency. We believe this concern is unfounded and may reflect a misunderstanding of our method. The reviewer writes:

My main technical concern remains that the unconditional 2D score model marginalizes over orientations and conformations and therefore does not explicitly enforce cross-view conformational consistency.

We agree that the 2D score model represents a marginal distribution over orientations and conformations and trivially does not explicitly encode cross-view conformational correspondences. The relevant point, however, is that cross-view geometric consistency is imposed during reconstruction rather than encoded explicitly in the score model. Projections at different orientations are coupled through a shared 3D volume: the anchor-view projection is constrained by the anchor image, while projections at sampled orientations are encouraged to be plausible under the score model. This enforces geometric consistency across views. Violations of the anchor constraint are explicitly penalized by the reconstruction objective (Equation 8).

Regarding the fixed-pose comparison, we refer to CryoBench (Section 5, p. 9), which explicitly states:

However, we note that fixed pose heterogeneous reconstruction is an unrealistic setting for cryo-EM of complex mixtures

Our work addresses the ab initio setting, in which particle poses are not supplied.

We agree that additional experiments would further strengthen the evidence for reliability and comparative performance. However, we already evaluate CryoSD on a broad and diverse set of highly challenging datasets. As the reviewer acknowledges, "The approach presented in this manuscript is innovative and exposes an orthogonal direction rarely explored for cryoEM reconstructions." Our contribution is to introduce this new concept and provide initial empirical evidence of its feasibility and advantages, demonstrating that it is more than a conceptual proposal. We do not claim that CryoSD has already reached the maturity of established software packages.

Reviewer_Lp7vOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CryoSD, which uses score distillation as a way of doing heterogeneous ab initio cryo-EM reconstruction without needing to do pose estimation. In this paper, a score-based diffusion model is pre-trained to denoise raw cryo-EM images. This pre-trained model is then used to do score-distillation-based 3D reconstruction.

Critical strengths

This paper’s biggest strength is the novel approach of using score distillation sampling (SDS) to do ab initio reconstruction.

Critical weaknesses

This paper has a few critical weaknesses:

(W1) I find it very hard to understand this paper: I found this paper difficult to understand for a few reasons: one, I felt that the main ideas of the paper were not clear. In particular, I think this paper would benefit from a section that explains the key ideas at a high level, with links to the sections that provide the details of each idea. Second, at times this paper is a wall of text that is not organized into paragraphs or sections. Examples include Section 3.2, which is one huge paragraph that takes up approximately one page, and the EMPIAR-10295 results paragraph, which also takes up approximately one page. I think separating these sections into paragraphs would help highlight the different ideas. For example, on Section 3.2, it may help to separate it into subsections or paragraphs regarding preprocessing, training the diffusion model, denoising, etc. Third, the notation in the math sections is confusing. There are a lot of variables in the paper and some are not defined until significantly after they are used in equations, such as ws,waw_s, w_a. Perhaps the authors should include a chart explaining all the variables in the Appendix.

(W2) The empirical support for this paper is weak: This weakness primarily comes from two sources: gaps in performance against the baselines when the comparison is fair, and unfair comparisons where CryoSD has a gap in performance.

On IgG-1D and IgG-RL, the proposed method’s performance lags behind cryoDRGN-AI’s. This can be seen from the FSC curves in Figure 3, where the red curve representing cryoDRGN-AI is always at or above the blue curve representing CryoSD. Additionally, quantitative metrics, such as FSC or FSC-AUC, are not reported, which would allow a quantitative comparison against competing methods that were not benchmarked in this paper but were benchmarked in other papers, such as cryoDRGN 2 (benchmarked in the CryoBench paper).

On the synthetic capsidHA dataset, the comparison to cryoDRGN-AI is unfair. CryoDRGN-AI assumes Gaussian white noise and could easily be adapted to handle the case of colored noise, for example by whitening the input data. CryoSD should be compared against a cryoDRGN-AI model that is adapted to colored, for example by whitening or changing the likelihood as was done for CryoSD, so we can assess how score distillation compares without the colored noise confound.

On EMPIAR-10295, I don’t think it’s fair to compare CryoSD against the result of a RELION ab initio job. A RELION ab initio job is just the initial stage of the RELION pipeline, so the results should be compared against the whole RELION pipeline, not its first stage. Second, both RELION and cryoDRGN-AI were only run on their default settings. Given that the input images represent large cages, I think it makes sense at a minimum to tune RELION’s mask diameter parameter and cryoDRGN-AI’s frequency marching parameters. Even better would be a more thorough attempt to tune the baselines on this dataset, perhaps as a grid search over each method’s most important hyperparameters. This is also possibly relevant for the capsidHA dataset as well. Additionally, it’s my experience that several runs of an ab initio algorithm may be needed to get one that converges, so multiple runs should be carried out for cryoDRGN-AI.

(W3) Missing comparisons: I believe that the paper should compare their diffusion model-based denoising to other denoisers, with one motivation being to establish whether the gap in performance with cryoDRGN-AI on IgG-1D and IgG-RL is due to poor denoising.

Questions for authors
  1. Can the authors explain how they get a denoising model with no access to clean groundtruth images?
  2. Can the authors explain how the anchors work?
  3. Can the authors explain where the score distillation part actually gets used?
  4. What are the quantitative metrics(FSC, per-image FSC-AUC) for CryoSD and cryoDRGN-AI in the IgG-1D and IgG-RL experiments? What are the FSC metrics for CryoSD and the baselines in the capsidHA and EMPIAR-10295?
  5. How does a cryoDRGN-AI model that is adapted for colored noise (either by whitening or changing the likelihood) compare to CryoSD in the capsidHA dataset?
  6. How does a tuned full RELION pipeline compare to CryoSD on EMPIAR-10295?
  7. How does a tuned cryoDRGN-AI model compare to CryoSD on EMPIAR-10295?
  8. How sensitive is CryoSD to scale σ\sigma, which is estimated by visual inspection?
  9. How does the denoising model compare to other cryo-EM denoising methods?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.