Understanding biomolecular function requires characterizing three-dimensional structure and conformational variability, for which cryogenic electron microscopy (cryo-EM) is a powerful tool. Cryo-EM produces noisy two-dimensional particle images in which molecular structure, particle pose, and imaging conditions are jointly encoded. Recovering biological information from these observations spans inverse problems including pose estimation, heterogeneity analysis, and reconstruction. These tasks are usually addressed independently despite shared protein structure and image formation, motivating a common representation across inverse problems. Since biologically meaningful differences may be spatially localized, we study self-distillation with complementary global and masked patch objectives. Because experimental dataset discrimination can reflect acquisition-specific cues, we develop controlled synthetic and experimental evaluations and find that structural discrimination benefits from dense spatial features. We introduce the EMber family of vision transformers, pre-trained on >50 million experimental particle images from 214 EMPIAR datasets together with simulated projections from 268,000 molecular assemblies. EMber improves upon existing methods at conformational and compositional discrimination by 14.85 and 14.26 percentage points, estimates conformational populations to within 1.59pp error with a linear probe, and produces particle assignments yielding reference-quality reconstructions in fewer epochs. We demonstrate how to learn representations that capture biological information for use across multiple cryo-EM inverse problems.
Figures & tables
Figure 1: Cryo-EM overview and shared inverse problems . (A) Proteins share similar structural features across molecular systems. (B) Cryo-EM images are generated with an image formation process involving projection, imaging effects, and measurement noise. (C) Downstream tasks attempt to recover different variables of the same underlying generative process, relying on overlapping structural and imaging information.
Figure 2: Augmentations can be unphysical . (A) 2D rotations of an image do not span all possible rotations of a 3D protein. (B) Image-space crops and blurs do not translate to the cryo-EM forward model.
Figure 3: Compositional heterogeneity classification on CAHRA experimental data. (A) Fine-tuning accuracy versus model FLOPs, with a supervised ResNet-18 reference. (B) Accuracy versus model FLOPs when linear probing frozen backbones, with a reference supervised ResNet-18. (C) Pooled linear/MLP readouts with a convolutional head that preserves spatial information.
Figure 4: Fine-tuning for Hsp90 conformation classification. We discretize Hsp90 conformations according to CMA and CMB chain-opening angles, each with 20 classes, and average accuracy across the two prediction heads (random chance: 5%). (A) Representative conformations and rendered particles at different SNRs. (B) Top-1 accuracy at SNR 0.1 versus model forward FLOPs. (C) Accuracy across SNRs. Model differences largely disappear at SNR 0.001, where all methods approach chance.
Figure 5
Figure 6: Particle grouping measured through reconstruction resolution. After each epoch of CryoDRGN-AI pose search, particles are clustered in the learned latent space and reconstructed using ground truth poses. (A) Tomotwin-8: Boxes show the spread in FSC=0.5 resolution across eight structures. The dashed line is an Oracle reconstruction obtained with ground truth group assignments. Notably, EMber is already near the Oracle after one epoch. EMber reaches near reference resolution within two epochs, and has less variation across the reconstructed clusters compared to baseline methods. (B) Hsp90: Reconstructions visualized after 2 epochs of CryoDRGN-AI pose search. Colors represent the signed distance to the reference surface (negative inside, positive outside).
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Dataset Composition. Visualization of statistics coming from our EMPIAR dataset and the OpenProteinSet containing assemblies. (A) Training set EMPIAR entries organized by coarse biological categories. (B) Pixel sizes of EMPIAR entries. At the same pixel dimensions, larger pixel sizes correspond to wider field of views. (C) Amino acid counts of synthetic and experimental datasets.
Parameter
Value or range
Sampling
Accelerating voltage
300kV
Fixed
Spherical aberration ( Cs )
2.7mm
Fixed
Amplitude contrast ratio
0.1
Fixed
Mean defocus
0.9 – 2.5μm
Uniform
Astigmatism
0 – 200A˚
Uniform
Astigmatism angle
0∘ – 180∘
Uniform
Appendix
Table 2: CTF and envelope parameters used for synthetic pre-training.
Figure 8: Example Dense Features. Experimental cryo-EM image, low-pass filtered for visibility (left), followed by the image features colored by PC1, PC2, and PC3, L2 vector norms of each patch, and, finally, the cosine distance between each patch’s features and those of the center patch. EMber produces dense, semantically meaningful features.
EMber
Reference models
Property
Small
Base
Large
Giant
DINOv2
DINOv3
Backbone
ViT-S/16
ViT-B/16
ViT-L/16
ViT-g/16
ViT-g/14
ViT-7B/16
Backbone parameters
21.4M
85.3M
304.2M
1133.7M
1.1B
6.7B
Backbone + heads parameters
233.2M
301.8M
523.9M
1359.7M
—
—
Transformer blocks
12
12
24
40
40
40
Input channels
1
1
1
1
3
3
Appendix
Table 3: Hyperparameter comparison between our models and the published DINOv2/v3.
Figure 9: Experimental data schedule. Visualization of the proportion of training batch that comes from experimental data. We pre-train on primarily synthetic images, before using a cosine schedule to mid-training where we sample using a dispersion equation.
Figure 10: Visualized CryoPPP Experimental Images. (Top) EMPIAR 10017 β -galactosidase and (Bottom) EMPIAR 10061 also β -galactosidase but bound to an inhibitor. Structurally similar, but present visually different, potentially due to experimental conditions.
Figure 11: Exploratory TRPV1 visualization. (A) Three randomly sampled query particles from each entry, resized to 384×384 and standardized separately. Intensities are displayed over [−3,3] standard deviations; this display clipping is not applied during evaluation. The images are not shown at a common physical scale. (B,C) PCA and UMAP fitted to normalized CLS features from all 29,000 feature-bank particles, using a ViT-S teacher checkpoint at step 79,999. Blue circles indicate EMPIAR-10005, orange triangles indicate EMPIAR-10059, and the other 27 entries are gray. UMAP uses 15 neighbors, a minimum distance of 0.1, and seed 42. This earlier checkpoint and CLS-only representation differ from those used for the reported k NN results.
Figure 12: Hsp90 Fine-tuning loss curves. We visualize the training loss across different learning rates when fine-tuning on the Hsp90 dataset.
Figure 13: Spike protein structures from the Spike MD simulation used for Spike population estimation task plotted as a function of the collective variables (principal component 1, principal component 2). The original distribution of structures is shown in (A), with the filtered structures in (B).
Figure 14: True Spike-Up population vs predicted Spike-Up population across a range of training dataset sizes used to fit the linear classifier, and stratified based on evaluation SNR level.
Figure 15: Fourier Shell Correlation plots of Hsp90 reconstructions. Rows correspond to different methods, and columns correspond to pose search epochs. The two curves represent the open and closed states. The dashed line indicates FSC= 0.5 .
Epoch 1
Epoch 2
Method
Closed
Open
Closed
Open
Oracle
4.48
4.50
-
-
Lookup Table
8.71
9.12
8.78
9.10
CNN Encoder
7.27
6.43
5.08
4.78
Cryo-IEF-B
6.43
5.46
7.42
6.08
CryoUNI-B
5.37
5.39
4.78
4.97
Appendix
Table 4: Class-wise reconstruction resolution at FSC =0.5 for Hsp90 at an SNR of 0.01, reported in Å. Reference class 0 corresponds to the closed state, and reference class 1 corresponds to the open state. Lower is better. The corresponding ground-truth resolutions are 4.48 Å (closed) and 4.50 Å (open).
Figure 16: Fourier Shell Correlation plots of Tomotwin-8 reconstructions. Rows correspond to different methods, and columns correspond to pose search epochs. Each color represents one of the 8 reconstructed structures. The dashed line indicates FSC= 0.5 .
Cryo-electron microscopy reconstructs biological structures from images of frozen specimens, from cellular volumes to atomic-resolution molecular models.
Single-particle cryo-electron microscopy (cryo-EM) has become a widely adopted technique for biomolecular structure determination. The conventional cryo-EM computational pipeline first combines many particle images to reconstruct an electrostatic potential (ESP) map and then fits an atomic model to the recovered map. Density reconstruction has high sample complexity, requiring large numbers of particle images and making structure determination high-cost and low-throughput, particularly for heterogeneous samples. Downstream atomic model building, in turn, becomes increasingly difficult as the resolution of the reconstructed map deteriorates. Protein structure prediction models provide strong sequence-derived priors on atomic structure, and experiment-guided approaches can use these priors to recover structures consistent with experimental measurements. Yet, in cryo-EM, such priors are typically integrated only after density reconstruction during atomic model fitting. We introduce Fold'EM, an inference-time framework that combines priors from protein generative models directly with cryo-EM particle images to determine atomic models from a small number of single particle images, bypassing both intermediate density reconstruction and downstream model building against the reconstructed map. Across synthetic and experimental cryo-EM datasets, Fold'EM recovers accurate atomic structures both with known particle orientations and in an ab-initio setting where orientations are inferred jointly with structure. In heterogeneous datasets, Fold'EM further resolves distinct conformational states from mixed particle populations without separately reconstructing a density map and building an atomic model for each state. We believe these results open new avenues for structure determination in the low-sample regime and for characterizing low-population conformational states directly from cryo-EM particles.
Advaith Maddipatla, Märt-Erik Mäeots, Marco Pegoraro +5
Institute of Science and Technology Austria · University of Zurich · Frankfurt Institute for Advanced Studies +2
Single-particle cryo-electron microscopy (cryo-EM) pose estimation is traditionally solved anew for each dataset, where iterative refinement is done from scratch while the estimator learns to store the molecule in its weights. In this work, we show that pose inference is a generalizable, specimen-agnostic operation when conditioned explicitly on a reference volume. We introduce ARCHER, an amortized contrastive classifier that models the pose posterior over a discrete rotation grid. Trained across a variety of protein structures, it operates zero-shot without retraining per structure. This transferability is grounded in Fourier-space information mechanics, where all specimen dependence is captured by the reference structure's power spectrum and spatial extent. ARCHER achieves a median angular error of 5.0° on 100 held-out test structures and 2.5° on experimental particles, matching dedicated estimators within 0.16 Å in 3D reconstruction. Crucially, downstream conformational signal is preserved. The leading conformational coordinate correlates at 0.97 with deposited benchmarks, faithfully reconstructing free-energy basins and mobile domains. These results overall demonstrate that cryo-EM pose estimation can be generalized across different structures.
Nhan D. Nguyen, Bao Pham
Pritzker School of Molecular Engineering, University of Chicago, Chicago, Illinois 60637, United States · Department of Computer Science, Rensselaer Polytechnic Institute, Troy, New York 12180, United States
Cryo-electron microscopy (cryo-EM) enables single-particle analysis of biological macromolecules under strict low-dose imaging conditions, but the resulting micrographs often exhibit extremely low signal-to-noise ratios and weak particle visibility. Image denoising is therefore an important preprocessing step for downstream cryo-EM analysis, including particle picking, 2D classification, and 3D reconstruction. Existing cryo-EM denoising methods are commonly trained with pixel-wise or Noise2Noise-style objectives, which can improve visual quality but do not explicitly account for structural consistency required by downstream analysis. In this work, we propose a score-based denoising framework for cryo-EM that learns the clean-data score to recover particle signals while better preserving structural information. Building on this formulation, we further introduce a target-guided variant that incorporates reference-density guidance to stabilize score learning under weak and ambiguous signal conditions. Rather than simply amplifying particle-like responses, our framework better suppresses structured low-frequency background, which improves particle--background separability for downstream analysis. Experiments on multiple cryo-EM datasets show that our score-based methods consistently improve downstream particle picking and produce more structure-consistent 3D reconstructions. Experiments on multiple cryo-EM datasets show that our methods improve downstream particle picking and produce more structure-consistent reconstructions.