Cryo-EM Image Processing and 3D Reconstruction: A Technical Reference for Veterinary Structural Virology

1. Introduction

Cryogenic electron microscopy (cryo-EM) has evolved into a cornerstone technique for determining the three-dimensional structures of macromolecular complexes at near-atomic resolution. The method enables structural biologists to visualize viral capsids, envelope glycoproteins, and host-pathogen interaction complexes without the need for crystallization. In veterinary virology, cryo-EM is increasingly applied to characterize emerging pathogens, inform vaccine design, and elucidate mechanisms of host range restriction. The image processing and 3D reconstruction pipeline remains the rate-limiting step, requiring sophisticated computational algorithms to extract high-resolution information from noisy, low-dose images [1].

This article provides a detailed review of the image processing and reconstruction workflow used in cryo-EM, with emphasis on the computational methods algorithmically linking raw micrographs to interpretable 3D density maps. The discussion is oriented toward veterinary applications, including the structural analysis of animal viruses such as Canine Parvovirus variants, Feline Leukemia Virus, and Highly Pathogenic Avian Influenza (H5N1).

2. Image Acquisition and Preprocessing

2.1 Data Collection and Motion Correction

Cryo-EM data are collected as movies comprising multiple frames, each exposed to a low electron dose to minimize radiation damage. Sample motion induced by electron beam interaction causes blurring across frames. Motion correction algorithms estimate per-frame shifts and apply alignment to produce a summed, dose-weighted micrograph.

The Unbend algorithm [2] models local beam-induced sample motion using a 3D spline interpolation. This approach corrects non-rigid deformations by fitting a smooth displacement field across the micrograph, preserving high-frequency information essential for subsequent reconstruction.

2.2 Contrast Transfer Function Estimation

The contrast transfer function (CTF) describes phase and amplitude modulation by the objective lens. Accurate CTF parameters (defocus, astigmatism, spherical aberration) must be estimated for each micrograph. WebCalEM [3] provides a browser-based tool for routine pixel size calibration, a critical step because incorrect pixel size propagates errors into 3D scaling. The tool uses power spectrum analysis to calibrate pixel size against known standards, reducing systematic errors in map interpretation.

2.3 Micrograph Quality Assessment

prismPYP [4] introduces a self-supervised learning framework that evaluates micrograph quality in both power-spectrum and image domains. By training a neural network without manual labels, it predicts a quality score reflecting ice thickness, contamination, and CTF fit, enabling automated filtering of poor-quality images.

3. Particle Picking

3.1 Supervised and Few-Shot Learning Approaches

Particle picking identifies individual macromolecular projections in micrographs. Traditional template-based methods require prior structural knowledge, which is often unavailable for novel viruses. Deep learning methods have largely supplanted these approaches.

CryoFSL [5] employs a few-shot learning paradigm that generalizes from few annotated examples. The model uses a siamese network architecture to compute similarity between a support set and query patches, achieving robust particle identification even when training data are sparse. This is particularly useful for veterinary pathogens where only a limited number of micrographs may be available.

CryoSIP [6] integrates semantic and instance segmentation through a collaborative picking network. It first generates semantic masks highlighting particle regions, then uses instance-level refinement to separate overlapping particles. This approach reduces false positives from ice contamination and aggregates.

3.2 Prompt-Guided Segmentation

CryoPromptSeg [7] adds a denoising module to a prompt-guided segmentation framework. The model accepts user prompts (e.g., bounding boxes or points) and produces refined segmentation masks. The integrated denoiser suppresses background noise while preserving particle edges, improving picking accuracy on low-contrast micrographs.

3.3 Viral Particle Specific Data Sets

CryoVirusDB [8] provides an annotated dataset of viral particles in cryo-EM micrographs, designed for training AI-based picking models. The dataset includes multiple virus morphologies (icosahedral, filamentous, enveloped) and is directly applicable to veterinary virology research.

4. Two-Dimensional Classification

After particle extraction, 2D classification groups projections into homogeneous classes representing different orientations and conformational states. This step assesses data quality and identifies preferred orientations.

AlignPCA-2D [9] reduces computational cost by applying principal component analysis (PCA) to the Euclidean vectors representing particle images. The PCA-reduced alignment space accelerates iterative classification while maintaining alignment accuracy. This method is compatible with standard maximum-likelihood classification frameworks.

5. Three-Dimensional Reconstruction

5.1 Ab Initio Reconstruction

When no initial model exists, ab initio algorithms generate a low-resolution 3D map directly from projection images. The subspace method of moments [10] leverages statistical moments of projection data to reconstruct the 3D density without iterative alignment. This approach is robust to noise and avoids convergence to local minima, making it suitable for veterinary viruses with unknown symmetry.

5.2 Orientation Determination and Refinement

Bayesian orientation determination [11] models the probability distribution of particle orientations given the observed projections. The Bayesian framework incorporates prior knowledge about the noise model and structural heterogeneity, enabling simultaneous refinement of orientation parameters and density map.

For helical assemblies such as nucleocapsids of Bovine Coronavirus, specialized pipelines are required. A dedicated cryo-EM processing pipeline for microtubules [12] uses helical symmetry to reduce the number of independent parameters, achieving near-atomic resolution for filamentous structures. The same principles apply to viral helical nucleocapsids.

5.3 Heterogeneity Analysis

Structural heterogeneity arises from flexibility, ligand binding, or conformational states. The Bayesian approach of Xu et al. [11] simultaneously reconstructs multiple 3D classes, assigning each particle to a conformational state. This method has been applied to map structural transitions in viral entry glycoproteins.

In cryo-electron tomography (cryo-ET), heterogeneity is addressed by subtomogram averaging. ICECREAM [13] uses equivariant neural networks to align subtomograms, preserving rotational symmetry and improving signal-to-noise ratio (SNR).

5.4 Resolution Assessment and Validation

Resolution is typically estimated by Fourier shell correlation (FSC) between two independently refined half-maps. The FSC threshold of 0.143 is widely used for near-atomic resolution claims. No single paper in the provided set specifically discusses general FSC methodology, but the practical applications in other references implicitly use these criteria.

6. Cryo-Electron Tomography Specific Workflows

6.1 Tomographic Data Acquisition and Alignment

Cryo-ET acquires a series of tilted projections from a single specimen region. Tilt series alignment is critical for accurate reconstruction. Traditional fiducial-based alignment uses gold beads, but marker-free approaches are gaining traction. Markerfree [14] implements GPU-accelerated feature matching to align tilt series without fiducials, improving throughput for lamella samples.

The Volta phase plate has been evaluated for improving tomogram alignment [15]. By increasing phase contrast at low defocus, the phase plate enhances alignment accuracy for weak-phase objects such as viral spikes.

6.2 Reconstruction and Denoising

Tomographic reconstruction from aligned tilt series produces a 3D volume. Composite sparse regularization [16] imposes sparsity constraints in both pixel and wavelet domains, reducing missing wedge artifacts and preserving fine details.

Denoising is essential for interpretability. A systematic comparison of deep learning denoising methods [17] evaluated architectures such as U-Net, noise2noise, and generative adversarial networks. The study proposed improvements in network design to balance artifact reduction with preservation of structural details.

The transition from filtering to denoising [18] reviews how classical Wiener filtering and modern neural network methods differ in visual interpretability. Neural network approaches generally outperform linear filters but can introduce hallucinated features if training data are mismatched to the target.

AreTomoLive [19] provides an automated pipeline for real-time tomographic reconstruction that includes comprehensive correction (motion, CTF, dose) and denoising. This pipeline enables online quality assessment during data collection, reducing the need for post-hoc reprocessing.

6.3 Subtomogram Averaging and Particle Analysis

TANGO [20] offers a complete framework for particle curation and analysis in cryo-ET. It includes tools for particle subtomogram extraction, classification, and averaging. The software integrates with established packages and provides quality metrics for each particle, facilitating reproducible subtomogram averaging.

For in situ single-particle cryo-EM, methods have been developed to extract particles directly from tomograms and perform structure determination without purification [21]. This is particularly valuable for studying viral assemblies within their native cellular environment, such as virions budding from infected cells.

7. Specialized Considerations for Veterinary Structural Virology

7.1 Beam-Induced Motion and Dose Sensitivity

Nucleic acids are more sensitive to electron radiation than proteins. The influence of total electron dose on the quality of nucleic acid potential maps was systematically examined [22]. For RNA viruses such as Avian Influenza, optimal dose fractionation and weighting are necessary to visualize the genomic RNA within capsids.

7.2 Calibration and Reproducibility

WebCalEM [3] addresses pixel size calibration, a frequent source of error in veterinary studies where accurate dimensions of viral spikes or capsid diameters are needed for vaccine development.

7.3 Integration with X-Ray Crystallography

A review of transitioning from X-ray crystallography to cryo-EM [23] discusses how complementary methods can accelerate structural studies of veterinary pathogens. For example, initial models from crystallography can serve as templates for cryo-EM refinement, reducing the computational burden of ab initio reconstruction.

8. Workflow Diagram

The following Mermaid diagram summarizes the cryo-EM image processing and 3D reconstruction pipeline for single-particle analysis.

graph TD
    A[Electron Microscope Movie Collection], > B[Motion Correction e.g., Unbend [<a href="#ref-2">2</a>]]
    B, > C[CTF Estimation e.g., WebCalEM [<a href="#ref-3">3</a>]]
    C, > D[Micrograph Quality Assessment e.g., prismPYP [<a href="#ref-4">4</a>]]
    D, > E{Particle Picking}
    E, > F[Few-Shot Learning e.g., CryoFSL [<a href="#ref-5">5</a>]]
    E, > G[Semantic-Instance Collaborative e.g., CryoSIP [<a href="#ref-6">6</a>]]
    E, > H[Prompt-Guided Segmentation e.g., CryoPromptSeg [<a href="#ref-7">7</a>]]
    F, > I[2D Classification e.g., AlignPCA-2D [<a href="#ref-9">9</a>]]
    G, > I
    H, > I
    I, > J[Ab Initio Reconstruction e.g., Subspace Method of Moments [<a href="#ref-10">10</a>]]
    J, > K[3D Refinement with Orientation Determination e.g., Bayesian [<a href="#ref-11">11</a>]]
    K, > L[Heterogeneity Analysis]
    L, > M[Final 3D Map]
    L, > N[Conformational States]
    M, > O[Resolution Validation FSC]
    N, > O

9. Emerging Technologies and Future Directions

9.1 In Situ Structural Biology

Cryo-FIB enables 3D imaging of biological specimens by milling thin lamellae from vitrified cells [24]. This technique allows visualization of viruses within host cells, bridging the gap between cellular and molecular resolution. For example, the structural organization of African swine fever virus assembly can be studied in situ.

9.2 Automated Workflows and Machine Learning

Automated cryo-EM combined with supervised machine learning has been used for reproducible characterization of extracellular vesicles and co-isolating particles [25]. Similar approaches are being adapted for viral particle identification in contaminated samples.

9.3 Integrated Tomography Pipelines

CryoFTM (mentioned in passing in several articles) and AreTomoLive [19] represent a trend toward fully automated, real-time processing. Future developments will likely integrate deep learning denoising and subtomogram averaging into a seamless online workflow.

9.4 High-Resolution for Veterinary Pathogens

The ongoing revolution in cryo-EM [1] continues to lower barriers for high-resolution structure determination. For veterinary virology, this means routine access to atomic models of viral proteins critical for antigenic characterization, such as the hemagglutinin of H5N1 or the spike protein of Feline Coronavirus.

References

[1] Subramaniam S, Kühlbrandt W, Henderson R. Cryo-EM: the revolution continues. IUCrJ. 2026. https://pubmed.ncbi.nlm.nih.gov/42258316/

[2] Kong L, Zottig X, Elferich J et al. Unbend, correction of local beam-induced sample motion in cryo-EM images using a 3D spline model. Elife. 2026. https://pubmed.ncbi.nlm.nih.gov/42059788/

[3] Dang L, Wang Z, Cho SHJ et al. WebCalEM: a browser-based tool for routine and accurate pixel size calibration in cryo-EM. bioRxiv. 2026. https://pubmed.ncbi.nlm.nih.gov/42244701/

[4] He L, Bartesaghi A. prismPYP: Power-spectrum and image domain learning for self-supervised micrograph evaluation. Structure. 2026. https://pubmed.ncbi.nlm.nih.gov/41864200/

[5] Poudel B, Gyawali R, Dhakal A et al. CryoFSL: an annotation-efficient, few-shot learning framework for robust protein particle picking in cryo-electron microscopy micrographs. Brief Bioinform. 2026. https://pubmed.ncbi.nlm.nih.gov/42248583/

[6] Deng Y, Wang S, Xiang M et al. CryoSIP: unleashing protein high-resolution Cryo-EM via semantic-instance collaborative picking. Brief Bioinform. 2026. https://pubmed.ncbi.nlm.nih.gov/41921195/

[7] Yang B, You Y, Jin L et al. CryoPromptSeg: prompt-guided segmentation with integrated denoising for cryo-EM particle picking. Bioinformatics. 2026. https://pubmed.ncbi.nlm.nih.gov/42172598/

[8] Gyawali R, Dhakal A, Wang L et al. CryoVirusDB: An Annotated Dataset for AI-Based Virus Particle Identification in Cryo-EM Micrographs. Viruses. 2026. https://pubmed.ncbi.nlm.nih.gov/41754567/

[9] Ramírez-Aportela E, Zarrabeitia OL, Fonseca YC et al. AlignPCA-2D: PCA-reduced Euclidean vector alignment for 2D classification in cryo-EM. Acta Crystallogr D Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/42257246/

[10] Hoskins J, Khoo Y, Mickelin O et al. Subspace Method of Moments for Ab Initio 3-D Single Particle Cryo-EM Reconstruction. SIAM J Imaging Sci. 2026. https://pubmed.ncbi.nlm.nih.gov/42220654/

[11] Xu S, Balanov A, Singer A et al. Bayesian perspective for orientation determination in cryo-EM with application to structural heterogeneity analysis. Acta Crystallogr D Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/41834536/

[12] Zhang D, Muñoz-Hernández H, Filipcik P et al. A cryo-EM processing pipeline for microtubules using CryoSPARC. Acta Crystallogr D Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/42044018/

[13] Kishore V, Debarnot V, Righetto RD et al. ICECREAM: high-fidelity equivariant cryo-electron tomography. Acta Crystallogr D Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/41892487/

[14] Xu Z, Liu Z, Niu T et al. Markerfree: GPU-accelerated marker-free alignment for improved cryo-ET reconstruction. Structure. 2026. https://pubmed.ncbi.nlm.nih.gov/41763201/

[15] Hutchings J, Ji D, Zheng S et al. Evaluating the Volta phase plate for improved tomogram alignment in cryo-electron tomography. IUCrJ. 2026. https://pubmed.ncbi.nlm.nih.gov/41954506/

[16] Yu C, Xu Z, Zeng Q et al. A variational framework with composite sparse regularization for cryo-electron tomography reconstruction. Bioinformatics. 2026. https://pubmed.ncbi.nlm.nih.gov/42191651/

[17] Jones HN, Deshmukh A, Pande K. Deep-learning methods for contrast enhancement and artifact reduction in cryo-electron tomography: a systematic analysis of the state of the art and proposed improvements. Acta Crystallogr D Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/41711797/

[18] Stoll T, Fäßler F. From filtering to denoising: Increasing visual interpretability of cryo-electron tomograms. Curr Opin Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/42190317/

[19] Peck A, Yu Y, Paraan M et al. AreTomoLive: automated reconstruction of comprehensively corrected and denoised cryo-electron tomograms in real time and at high throughput. Nat Methods. 2026. https://pubmed.ncbi.nlm.nih.gov/42185541/

[20] Schreiber M, Turoňová B. TANGO: Analysis and curation of particles in cryo-electron tomography. Nat Commun. 2026. https://pubmed.ncbi.nlm.nih.gov/41667456/

[21] Zhang X, Huang G, Sui SF. In situ single-particle Cryo-EM methods: Principle and applications. Curr Opin Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/42134001/

[22] Ważny G, Jaciuk M, Indyka P et al. Influence of total electron dose on the quality of nucleic acids potential maps in Cryo-EM. Ultramicroscopy. 2026. https://pubmed.ncbi.nlm.nih.gov/41886978/

[23] Zafar H, Malone KL, Singh AK et al. Breaking barriers: transitioning from X-ray crystallography to cryo-EM for structural studies. Acta Crystallogr D Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/41709867/

[24] Son S, Lee S, Kim J et al. Cryo-FIB-enabled 3D imaging of biological specimens. Appl Microsc. 2026. https://pubmed.ncbi.nlm.nih.gov/41991878/

[25] Enciso-Martinez A, Faas FGA, de Jong AWM et al. Automated Cryo-EM and Supervised Machine Learning Enable Reproducible Characterization of Extracellular Vesicles and Co-Isolating Particles. J Extracell Vesicles. 2026. https://pubmed.ncbi.nlm.nih.gov/41979060/

[26] Sommerhage S, Lammert J, Camino JD et al. Cryo-EM structure determination of amyloid fibrils: Methodology, insights, and practical tools. Methods Enzymol. 2026. https://pubmed.ncbi.nlm.nih.gov/42055641/

[27] Iragavarapu AG, Artemchuk O, Bobe D et al. A robust workflow for 3D imaging of human mitochondria using cryo-electron tomography. bioRxiv. 2026. https://pubmed.ncbi.nlm.nih.gov/42039535/

[28] Zhu Y, Zhang Y, Li S et al. Rising Stars: From the Development of Cryo-Electron Tomography to In Situ Structural Biology Research. J Mol Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/41796651/

[29] Wiedemann S, Fabian Z, Soltanolkotabi M et al. ProPicker: Promptable segmentation for particle picking in cryogenic electron tomography. J Struct Biol. 2026. https://pubmed.ncbi.nlm.nih.gov/41763621/