acceptodds
Under review as a conference paper at ICLR 2027

Separating signal from noise: a self-distillation approach for amortized heterogeneous cryo-EM reconstruction

Abstract

Cryogenic electron microscopy (cryo-EM) is a powerful method for visualizing heterogeneous biomolecular structures at high resolution, but the growing scale of cryo-EM datasets creates substantial computational challenges for 3D reconstruction. Autoencoder-based heterogeneous reconstruction methods offer a potential solution through amortized inference: once trained, an encoder could infer the conformations of new particle images with a single forward pass. In practice, however, encoders of existing methods overfit to the extremely noisy images used for training and fail to generalize to unseen particles. Here, we introduce Cryo-No-Overfit (CryoNOO), a training paradigm for learning noise-invariant representations that generalize to unseen cryo-EM images. CryoNOO leverages the reconstruction model itself as a denoiser, reconstructing and reprojecting training particles to generate denoised views that are augmented with independent noise realizations for self-supervised learning. Across synthetic and experimental datasets, we show that CryoNOO enables large-scale amortized reconstruction from a subset of particles, generalizes to conformations absent during training, and supports accurate conformation inference on unseen particles.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
cryo-EM, cryogenic electron microscopy, generalization, distillation, self-supervised learning, self-distillation, representation learning

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

This paper addresses encoder overfitting in autoencoder-based heterogeneous cryo-EM reconstruction. Methods such as cryoDRGN could in principle amortize conformation inference, but their encoders overfit to noise in the training images and generalize poorly to unseen particles, preventing training on a subset of a large dataset and inferring the rest. The authors propose CryoNOO, a training objective added to cryoDRGN. The reconstruction model’s own rendered projection of each training particle serves as a denoised view. Fresh Gaussian noise is added to this view, and a VICReg loss encourages the encoder to produce consistent latents for the original and re-noised images. The added components are used only during training, so inference is unchanged. The paper evaluates the method on CryoBench-derived synthetic datasets and on CAHRA data. It reports three finding: (i) better reconstruction of held-out particles, including training on 10% of a 1M particle dataset and inferring the remainder; (ii) placement of particles from a held-out conformation at the expected location in latent space; and (iii) conformation inference from the learned latents that is competitive with cryoSBI.

Critical strengths
  • A well-motivated, simple fix for a documented failure mode, with the right controls. Encoder memorization in cryoDRGN-style models is a known bottleneck, and the CryoNOO’s fix requires only an expander head and one additional loss on top of an unmodified cryoDRGN (Sec. 3.2). The expander is discarded after training and inference is identical to cryoDRGN. Crucially, the ablations rule out the trivial explanations: adding noise to the original images, invariance-only training, and SimSiam all underperform the full objective (Table 3), so the gain is attributable to the combination of model-generated denoised views and the full VICReg objective.

  • A practically meaningful use case demonstrated at scale, with consistent gains across settings. The 100K to 1M amortization experiment (Table 1) matches a real workflow, gains over cryoDRGN-100K appear on all five CryoBench datasets (Table 6) and on experimental CAHRA data including HQ to LQ transfer (Table 7), and the latent space analyses (Table 9-11) corroborate the mechanism rather than just the endpoint.

Critical weaknesses
  1. The gains are not cleanly attributed to the claimed mechanism, and no variability is reported.

    • Sec. 3.2 states gradients are not stopped through X̂. VICReg therefore also updates the decoder, so the method is not a teacher-student set up and encoder-side and decoder-side effects are entangled. No stop-gradient ablation is given.
    • The ablations do not isolate the paper’s key insight “reconstruction model as denoiser”, namely that the decoder-rendered projection serves as a denoised view. The direct control is to keep the full VICReg objective but construct views from the raw image instead (X vs. X + ε, no rendering). Neither existing variant in Table 3 provides this control. “CryoDRGN + Noise Aug” (0.328) appears to apply noise augmentation without VICReg (Sec. 4.4 is not explicit), which removes both components at once. “Only Invariance” (0.325, identical to the cryoDRGN baseline) ablates VICReg’s regularizers while keeping X̂.
    • Table 7 is hard to reconcile with an encoder-generalization claim. On CAHRA HQ→HQ, CryoNOO improves strongly on training particles. CryoDRGN’s test metrics are at or above its train metrics on ADH, Aldolase and PKM2, so there is no generalization gap to close. Those single-protein sets also have no conformational heterogeneity. Part of the gains may come from other pipeline differences or from a decoder-side regularization effect.
  2. The headline practical claims outrun the evidence.

    • Large-scale amortization is shown only on synthetic data, at SNR 0.1 (App. A), which is less noisy than the SNR 0.01 setting that also appears in Table 4. On the one heterogeneous real dataset (“Combined”), test gains are marginal (Table 7a: FSC 0.335→0.356, CD 1.132→1.120, vIoU 0.530→0.538), and CD is slightly worse in HQ→LQ (1.074→1.095, Table 7b).
    • HQ→LQ is described as validating the subset-train/infer-rest pipeline (App. C.1). But LQ and HQ particles come from the first and last 30 frames of each movie (App. C.1.1). If these are the same physical particles, this tests dose/noise robustness rather than generalization to unseen particles.
    • Unseen-conformation generalization (Sec. 4.2, Fig. 3) is shown only qualitatively, for three held-out states on one dataset, and each held-out state is directly flanked by training states, making this a mild interpolation test. The visual criterion is also not matched: the “expected location” boxes for CryoNOO cover roughly a quarter of the latent disk, whereas those for cryoDRGN are small patches on the ring. Moreover, cryoDRGN already generalizes poorly to unseen particles of seen conformations on IgG-1D (test CD 4.570, Table 6), so the scattering of held-out-conformation particles in Fig. 3 does not by itself isolate the unseen-conformation effect. The 2D projection method for the 8-D latent is not stated.
    • The real-data noise handling is heuristic. A factor of 10× over the estimated σ is recommended (App. A), whereas the robustness claim in Sec. 4.4 rests on 0.5×–2× on synthetic data with known σ.
Questions for authors
  • No table reports variability across runs, and per-image FSC is computed from one image per conformation (App. E). Please report mean ± std over at least three seeds for the main comparisons (Tables 1, 2, 6, 7), and the test-set size for Table 2, so that small differences (e.g., IgG-RL and Spike-MD in Table 6; CryoNOO vs. cryoSBI in Table 2) can be assessed.
  • Table 3 is evaluated on test per-image FSC, and its chosen configuration matches the IgG-1D test result in Table 6. The CAHRA noise factor of 10 was also selected based on observed performance (App. A). Were these choices made on the test split? If so, please re-select them on a held-out validation split and report the resulting test performance.
  • Add controls on synthetic data: (i) VICReg between two noisy copies of the real image with no decoder-rendered view; (ii) a stop-gradient on X̂; (iii) an oracle using clean ground-truth projections plus fresh noise. Please also report latent variance across noise draws of the same clean image for cryoDRGN vs CryoNOO.
  • Are the HQ and LQ particles in HQ→LQ the same physical particles? If so, please evaluate on genuinely held-out particles.
Reviewer_std9OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CryoNOO, a self-distillation approach that extends cryoDRGN with noise-invariant representation learning. It uses reconstructed projections with independent noise augmentations and a VICReg objective to improve amortized inference on unseen particle images.

Critical strengths
  • The method addresses a relevant limitation of amortized cryo-EM reconstruction through a simple training objective without increasing the inference network size.
  • Experiments cover multiple synthetic and experimental datasets, with useful ablations of noise augmentation and the self-supervised objective.
Critical weaknesses
  • Insufficient validation of unseen-conformation generalization. The experiment holds out a single state while retaining neighboring states along a densely sampled trajectory, primarily testing local interpolation. Latent-space visualizations without quantitative reconstruction assessment of the held-out states provide limited support for broader generalization claims.
  • Unclear baseline inference settings. The claimed pose-free inference protocol does not sufficiently explain test-time image centering or the use of translation parameters. Normalization and CTF preprocessing also need clarification to establish that CryoNOO and cryoDRGN receive equivalent inputs and that preprocessing differences do not contribute to the observed performance gap.
  • Insufficient statistical validation. The stated FSC evaluation samples one particle per conformation, but uncertainty across repeated particle sampling or independent training runs is not reported. Consequently, the robustness of the improvements, particularly the smaller gains on some datasets, remains unclear.
Questions for authors
  1. Can the authors evaluate held-out contiguous ranges of conformations and report quantitative reconstruction accuracy for these states, in addition to latent-space visualizations?
  2. Can the authors clarify the preprocessing and use of test-time translation and CTF parameters for both methods, and report variability across multiple random seeds and repeated test-particle sampling?
Reviewer_6WkcOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper addresses the issue of encoder overfitting in cryoDRGN which is a method for 3D reconstruction of heterogeneous cryo-electron microscopy (cryo-EM) data based on autoencoders. Because the encoder tends to fit the noise present in the training particles and fails to generalize to unseen particles, the amortized inference approach often performs poorly in practical applications. The proposed method, CryoNOO, leverages a pre-trained cryoDRGN model, utilizing the model itself as a denoiser. Experimental results demonstrate that training on a subset of a large synthetic dataset and generalizing to the remaining data yields better performance than training a standard cryoDRGN model on the same subset, while incurring significantly lower computational costs. Furthermore, conformations not included in the training set are mapped to appropriate locations within the latent space. On the IgG-1D dataset, latent-space-based k-nearest neighbor (kNN) and centroid-based detection methods outperform both cryoSBI and cryoDRGN in conformational inference tasks. The paper also validates key design choices through ablation studies. The primary contributions of this work include enabling large-scale amortized reconstruction using only a subset of data for training, ensuring the latent space generalizes to previously unseen conformations, and facilitating accurate conformational inference for new particles.

Critical strengths
  • Clear writing and motivation. The paper is well written. The problem is framed as one mechanism, the encoder fitting noise in the training images, and the method follows directly from that framing. Fig. 1 together with Sec. 3.2 gives a clear picture of the training pipeline, and each experiment in Sec. 4 is tied to one claim.
  • The design is simple and plug-and-play. CryoNOO adds one training loss and a projection head that is discarded after training, so inference is unchanged from cryoDRGN. This makes it easy to adopt in existing pipelines and the cost is moderate.
  • Performance improvements remain consistent on the held-out set. Across all five CryoBench datasets and various CAHRA configurations, the FSC-AUC metric consistently outperforms that of cryoDRGN.
Critical weaknesses
  • The key components were not isolated for individual examination. The core concept involves applying VICReg to re-projected data generated by the model itself that has been subjected to additional noise. Table 3 compares this approach against several baselines, including scenarios that rely solely on standard noise augmentation, use only invariance constraints, employ SimSiam, or omit the expander. However, the paper does not provide a comparison with a setup where the full VICReg framework is applied directly to the original images, bypassing the re-projection step. Without this control condition, it is impossible to disentangle the gains derived from self-distillation from the effects of standard noise-augmented VICReg.

  • There is a lack of assessment regarding variability across multiple runs. All results presented in the main text are derived from single runs, with no information provided on random seeds or error bars. This raises concerns on two levels. First, the proposed method combines VAE loss and VICReg loss during training and operates on views subjected to stochastic re-noising. Consequently, both the noise sampling process and the model initialization introduce randomness. Second, many of the conclusions drawn in the paper are based on marginal differences. For instance, in Table 3, the test FSC-AUC values ​​for various configurations fall within a narrow range where the selected default configuration is not even the best-performing one, and some design choices are justified based on performance gaps of only 0.01 to 0.02. Without data from multiple repeated runs, it is hard for the reader to discern which design choices have a substantive impact and which differences merely stem from fluctuations attributable to the random seed.

Questions for authors

Is it possible to evaluate the performance of VICReg by using the same views, expander, and loss weights on the original images after re-applying noise but omitting the reprojection step? Given that completing one epoch for IgG-1D on a single H100 GPU takes only about 1 to 1.5 minutes, the cost of a full run should be relatively low. Could you please provide the mean and standard deviation calculated from three runs with different random seeds for the results presented in Table 3, as well as for the rows in Table 1 corresponding to the 100K dataset size? This would help determine whether the differences observed in the ablation studies and the comparisons at the 100K data scale are statistically significant, rather than merely reflecting random fluctuations across multiple runs.

Comments on the paper

1 comment

Thank you for the response and for running the controls. The new experiments address my main concerns well. The no-reprojection and matched-noise variants isolate the contribution of the rendered view, the oracle shows little headroom is left in this setting, and the stop-gradient variant both resolves the encoder/decoder entanglement question and improves the method. The multi-seed results in Tables B, C, and F show the main gaps are well outside run-to-run variability. The noise-factor answer is also satisfying: the 10x factor was selected on two separate experimental datasets rather than on the CAHRA evaluation data, and the Aldolase sweep shows a flat plateau from 10 to 50. The clarification that HQ and LQ particles are not the same physical particles validates HQ->LQ as a genuine unseen-particle test. Overall the rebuttal strengthens the paper, and I will raise my rating from weak reject to weak accept.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.