Resolving structural heterogeneity in situ through cryogenic electron tomography

Abstract

Cryogenic electron tomography (cryoET) has emerged as a powerful tool for studying the structural heterogeneity of proteins and their complexes, offering insights into macromolecular dynamics directly within cells. Driven by recent computational advances, including powerful machine learning frameworks, researchers can now resolve both discrete structural states and continuous conformational changes from 3D subtomograms and stacks of 2D particle-images acquired across tilt-series. In this review, we survey recent innovations in particle classification and heterogeneous 3D reconstruction methods, focusing specifically on the relative merits of workflows that operate on reconstructed 3D subtomogram volumes compared to those using extracted 2D particle-images. We additionally highlight how these methods have provided specific biological insights into the organization, dynamics, and structural variability of cellular components. Finally, we advocate for the development of benchmarking datasets collected in vitro and in situ to enable a more objective comparison of existent and emerging methods for particle classification and heterogeneous 3D reconstruction.

Introduction.

Since David DeRosier and Aaron Klug highlighted the promise of combining multiple tilted views to reconstruct a three-dimensional density map of a structure^1,2, the field of electron tomography has made steady progress toward visualizing macromolecular complexes in three dimensions – first using purified samples and later with vitrified cells. Indeed, early studies, including pioneering work by Baumeister and colleagues, demonstrated the potential of cryogenic electron tomography (cryoET) for in situ structural analysis by visualizing the coarse-grained architecture of a eukaryotic cell and resolving large, abundant cytosolic complexes, such as the ribosome and proteasome³. Over the subsequent two decades, cryoET has provided insights into cellular organization and architecture from a growing list of systems including prokaryotic, archaeal, and eukaryotic cells, as well as isolated organelles and viruses^4-7. In parallel, ongoing improvements in subtomogram averaging (STA) methods, which align and average multiple copies of a macromolecule extracted as three-dimensional subvolumes from tomograms (i.e., subtomograms), have enabled researchers to visualize complexes in situ at near-atomic resolution and to investigate their conformational and functional states.

Despite these successes, cryoET faces inherent challenges that require innovative computational solutions. Compared to single-particle analysis (SPA), the use of tilt-series data in cryoET presents both opportunities and challenges for the reconstruction of structurally heterogeneous molecules. In one regard, by providing multiple views of individual particles, tilt-series image acquisition increases projection angle sampling, which is especially beneficial for asymmetric particles that adopt preferred orientations, and it helps distinguish particles positioned atop one another along the imaging axis. Simultaneously, it introduces challenges such as reduced signal relative to noise (often described as the signal-to-noise ratio, or SNR) within individual images resulting from the lower electron dose per image, a thicker imaging path at higher tilt angles, and increased radiation damage for images late in the acquisition series. Additionally, tilting the sample introduces defocus gradients across the images that must be considered. Finally, when performing tomogram reconstruction, errors in coarse tilt alignment estimation can limit the resolution of reconstructed tomograms, and often necessitate extensive computational refinement of alignments to achieve high-resolution reconstructions of the macromolecular complexes of interest^8-12.

Advances in computational methods^11-15 now effectively address many of the inherent challenges in cryoET, allowing for the reconstruction of both discrete and continuous structural ensembles of protein complexes spanning scales from ~200 kDa nucleosomes¹⁶ to the ~300 kDa Influenza A ribonucleoprotein complex¹⁷ to multi-megadalton photosynthetic complexes¹⁸. The Electron Microscopy Data Bank (EMDB) is growing exponentially with such work, both in the number of studies published and in the number of structures resolved within a given study, highlighting how researchers are increasingly leveraging the rich data cryoET provides (Figure 1A). Indeed, through their recent in situ analysis of the eukaryotic chaperonin TRiC, which folds nearly 10% of the cellular proteome through dynamic ATP-driven structural transitions, the Beck and Frydman groups showcased this growing capability¹⁹. Specifically, they employed RELION’s 3D-classification method to resolve multiple conformational states of TRiC directly in cells, effectively differentiating TRiC assemblies with one or two prefoldin (PFD) domains observed, and they resolved a conformational change in the chaperonin that closes the chamber, thereby encapsulating its client protein (Figure 1B-C).

Figure 1. Capacity and efficacy of in situ structural analyses are growing rapidly.

(A) Meta-analysis of Electron Microscopy Data Bank (EMDB) deposited structures resolved by cryoET. Each point represents a unique study (DOI) associated with one or more structures produced by subtomogram averaging (STA) that was deposited in the EMDB between 2007 and 2024. The x-axis is binned into three-year intervals, each labeled with the total number of STA structures deposited during that period. The y-axis indicates the number of STA structures deposited per DOI, serving as a proxy for dataset throughput and/or reconstruction diversity within a single study. Points are color-coded by the median resolution of the STA reconstructions associated with each DOI, following color scale (right). Note time-dependent increase in total number of structures, average number of structures per study, and resolution. These data were retrieved from the EMDB in January 2025 and includes all entries annotated under the method "subtomogram averaging." (B) 3D-classification and reconstruction of the disparate states of the TRiC chaperonin complex, highlighting conformational and compositional diversity, performed in RELION. The core TRiC structure is colored grey, with variable regions, including a prefoldin domain (PFD) and inferred client proteins, highlighted in yellow. (C) Conformational states depicted in B, mapped to their original positions within the tomogram, revealing the spatial organization and diversity of TRiC processing states in the native cellular environment. Arrows indicate apparent clusters of complexes. Panels B-C were adapted from Xing et al., 2025.

Structure determination using 2D particle-image stacks or 3D subtomogram volumes.

Methods for heterogeneous reconstruction and classification in cryoET can be broadly categorized along two key axes (Table 1): 1) those that operate on stacks of 2D particle-images versus 3D subtomogram-volumes; and 2) those typically used for discrete classification versus those allowing for continuous classification. Whereas the field has traditionally used 3D subtomograms for downstream STA, it increasingly uses workflows that operate on 2D particle images extracted directly from the tilt-images. Indeed, early tomographic methods enabled the extraction of 3D sub-volumes from reconstructed tomograms and, as this approach is intuitive, it found widespread use as practitioners worked to determine coarse molecular structures and to resolve conformational states within those structures using classification-based tools. Indeed, discrete classification of 3D subtomograms was used in analyzing the TRiC chaperonin¹⁹, depicted in Figure 1. However, the use of such 3D subtomograms introduces key challenges, including a requirement to interpolate missing data from the limited tilt range - both between tilts and for the entire swath of the ‘missing wedge’ - and the managing the increased computational and storage demands of the interpolated three dimensional data relative to the 2D tilt-image data^8,20-22. As such, the field has recently shifted away from 3D subtomograms and towards using 2D particle stacks, an approach that reduces storage and computational costs while enabling higher-resolution reconstructions. By operating directly on tilt-series images, many modern tools allow for fine-grained, per-particle refinement of tilt geometry, motion, and CTF parameters during reconstruction^{9,11,14,15,23} - capabilities that are difficult to implement with interpolated 3D subtomograms. Related tools to resolve structural heterogeneity using such 2D data have also been developed and have proven highly efficacious.

Table 1. Overview of particle classification and heterogeneous reconstruction tools.

Categorization of methods for cryoET heterogeneity analysis based on the key attributes noted in columns, with reference to each method additionally provided.

Tool	Classification	Data Type	Approach to Heterogeneity	Priors	Reference
RELION 5.0	Discrete	2D	Bayesian inference	Pose, CTF, class #	Burt et al. 2024
NextPYP	Discrete	2D	Linear decomposition	Pose, CTF, class #	Liu et al. 2023
RELION 3.1–4.0	Discrete	3D	Bayesian inference	Pose, CTF, class #	Bharat and Scheres 2016
STOPGAP	Discrete	3D	Linear decomposition	Pose, CTF, class #	Wan et al. 2024
EMAN2	Discrete	3D	Multi-model refinement (and others)	Pose, CTF, initial models	Galaz-Montoya et al. 2015
tomoDRGN	Continuous	2D	Machine learning	Pose, CTF	Powell and Davis 2024
CryoDRGN-ET	Continuous	2D	Machine learning	Pose, CTF	Ragnan et al. 2024
emClarity	Continuous	2D	Linear decomposition	Pose, CTF	Himes and Zhang 2018
OPUS-TOMO	Continuous	3D	Machine learning	Pose, CTF	Luo et al. 2024
Dynamo	Continuous	3D	Linear decomposition	Pose, CTF	Castaño-Díez et al. 2012
PEET	Continuous	3D	Linear decomposition	Pose, CTF	Heumann et al. 2011
MDTOMO	Continuous	3D	Normal mode analysis of MD simulation	Pose, CTF, atomic model	Vuillemot et al. 2023
TomoFlow	Continuous	3D	Dense optical flow	Pose, CTF	Harastani et al. 2022

Discrete vs continuous classification.

The traditional classification methods, such as those implemented in RELION¹² and EMAN2²⁴, are typically used to group particles into a fixed number of predefined classes that, in principle, each represent a specific structural state. This approach is ideal for systems where a limited number of structural states are expected, or number of particles is limited. By concentrating particles into a relatively small number of classes, classification tends to maximize the number of particles in each class, thereby improving the signal-to-noise ratio of structurally homogeneous regions of the map and ultimately enabling high-resolution reconstructions for the most abundant states. However, forcing particles into discrete classes inherently limits the exploration of structural diversity, and it can result in rare or unexpected states being averaged into broader categories, potentially obscuring biologically significant heterogeneity.

In contrast, continuous classification methods map particles into a low-dimensional ‘latent’ space, or manifold, which allows for the representation of a conformational continuum. This approach is particularly advantageous when analyzing systems with complex and variable structural landscapes bearing an unknown, or unknowable, number of structural states. Indeed, by forgoing the need to initially restrict particles to discrete categories, continuous classification has the potential to capture a broader range of conformations, and to provide a more comprehensive view of the structural landscape in vitro and in situ^25-27, albeit often at lower resolution. We note that this continuous space can, of course, be used to partition particles into discrete sets post hoc, providing sets of particles that can often be resolved to high-resolution with dedicated refinement and reconstruction packages, such as M¹¹. It is additionally notable that methods for continuous classification typically use a vastly increased parameter space, and often use various forms of regularization in particle classification to limit overfitting to the high levels of noise inherent in cryoET.

Methods operating on 3D subtomograms vs 2D particle stacks.

The introduction of STA workflows in RELION 3.1⁸ marked a significant milestone for the field of cryoET, as it extended the SPA workflow, which was familiar to many researchers, to subtomogram volumes. Specifically, individual 3D subtomograms extracted from a tomogram could be sorted amongst a fixed number (commonly ~3-6) of discrete classes. Once assigned, individual classes could be further refined often resulting in higher resolution reconstructions. Additionally, integration of such subtomograms into standard RELION workflows allowed for the application of real space masks, allowing users to focus the classification on specific regions of interest, which has proven valuable in identifying distinct structural states across an array of macromolecules^18,28-37. Notably, this classification and refinement strategy can be performed iteratively and hierarchically, allowing one to better reveal subtle but important structural differences^38-43 using a relatively small number of classes in each round of classification, which is vital both for computational efficiency and to increase the effective number of particles assigned to each class.

Relatedly, STOPGAP⁴⁴ performs STA and discrete classification within the 3D subtomogram space. It employs a stochastic hill-climbing algorithm complemented by simulated annealing⁴⁵ during early alignment stages that, in combination, aim to avoid local minima during orientation and class assignment. Classification typically begins with a set of initial reference volumes, either provided manually or generated by randomly partitioning particles into subsets. These references serve as the starting point for iterative classification and refinement, where each round of classification involves aligning subtomograms to current references. Due to the stochastic nature of these methods, the classification can be run repeatedly, and STOPGAP can identify and retain particles that consistently classify into the same group while excluding inconsistently assigned particles. This process, known as consistency cleaning, ensures that only the most reliable particles are included in the final classes. Combined, it is thought that these features make STOPGAP particularly effective at classification, and it has been widely used^46-49.

More recently, the aforementioned advantages of working directly with 2D tilt-series images have become more widely appreciated, and newly developed methods have increasingly adopted this approach. RELION 5.0¹⁵ allows for a full 2D particle-stack workflow through its “extract subtomos” job that combines particle positions and tomogram alignment parameters to crop regions around particles in the tilt-series micrographs, thereby generating stacks of 2D particle-images. To address challenges related to the wide range of image quality often observed across a stack of 2D tilt-images, RELION 5.0 includes options to exclude images above a user-defined electron dose threshold and to filter particles that are not visible in a minimum number of tilt images. These features help reduce the impact of variable electron dose and missing views. Once processed, these stacks can be used in the traditional RELION pipeline, including ab initio reconstructions, refinement, discrete 3D classification, and others^15,50-52.

NextPYP¹⁴ further builds on the 2D particle-image concept, offering a scalable, end-to-end framework for cryoET data analysis that eliminates the need to generate large tomograms or subtomogram volumes. Instead, NextPYP processes raw, unaligned tilt series images directly, extracting particle images on-the-fly as needed and thus bypassing the need for intermediate particle stacks. This on-the-fly extraction approach further reduces the storage footprint of cryoET datasets and NextPYP’s seamless integration of tools for motion correction, CTF estimation, particle picking, 3D reconstruction, and analysis of structural heterogeneity through discrete and continuous (see below) methods significantly accelerates data processing.

Continuous classification – linear decomposition.

Whereas discrete classification methods have been foundational in the development of STA, they are inherently misaligned with macromolecular complexes that undergo continuous conformational changes, or those that adopt a large number of discrete states. For researchers aiming to resolve continuous conformational transitions or explore more complex structural landscapes, continuous classification and reconstruction strategies are essential^53,54. These approaches can be broadly categorized into traditional linear decomposition methods and more advanced non-linear algorithms, including machine learning-based methods, with variants of each developed to operate on 3D subtomograms or on stacks of 2D particle-images.

Principal component analysis (PCA) and singular value decomposition (SVD) are foundational linear decomposition techniques that have been applied in cryoET data processing suites including Dynamo⁵⁵, i3⁵⁶, and PEET⁵⁷. These techniques decompose sets of extracted 3D volumes into orthogonal components, enabling the identification of dominant modes of variability. By analyzing these modes, researchers can map particles onto a continuous conformational space, providing a low-dimensional representation of heterogeneity. Many of the software packages using these methods interlace reconstructions and classifications by calculating the covariance matrix of aligned particles during each iteration. PCA on this matrix projects the data into a low-dimensional space defined by the dominant modes of structural variance. This projection facilitates the visualization and exploration of particle distributions, potentially highlighting groupings related to underlying structural differences, which can be discretely clustered or sampled dynamically^10,58-61. Additionally, Dynamo and PEET allow users to specify spatial masks, enabling the classification of specific regions that exhibit structural variability.

Whereas most approaches rely on subtomogram averaging and iterative classification to boost the signal-to-noise ratio, methods like MDTOMO²⁵ and HEMNMA-3D⁶² adopt a different strategy by operating directly on individual subtomograms, with the MDTOMO method having proved particularly powerful in analyzing conformational motions in the SARS-CoV-2 spike protein²⁵. To mitigate the impact of the inherently high noise for single subtomograms, these methods integrate physics-based priors, such as normal mode analysis (NMA) or molecular dynamics (MD) simulations, to guide flexible fitting of atomic models into subtomograms. This enables the recovery of continuous heterogeneity at the atomic level, constrained by energetically and structurally plausible motions. The resulting deformation parameters are then projected into a low-dimensional space using PCA, as described above. This framework enables researchers to recover prominent structural transitions in terms of modeled atomic displacements, and to cluster or average density maps at specific positions along the energy landscape, offering a detailed and interpretable view of continuous heterogeneity.

Complementing the linear decomposition approaches applied directly to ensembles of 3D subtomograms, the emClarity approach developed by the Zhang group²³ incorporates biologically meaningful priors into the analysis, but in the 2D tilt-series space, rather than in the 3D subtomogram space. By applying band-pass filters, emClarity computes intervoxel correlations at biologically relevant length scales, targeting structural features such as ~10 Å for α-helices, ~20 Å for RNA helices, and ~40 Å for protein domains. SVD is then applied independently at each scale, and the resulting dominant components are concatenated into a feature vector for downstream clustering and classification. Additionally, emClarity incorporates particle-based tilt-series geometry refinement, similar to the refinement methods described in Warp¹³ and M¹¹. This iterative process refines both particle alignments and the geometric parameters of the tilt series using local, patch-based optimization, allowing the correction of tilt deformations while using regularization to limit overfitting. These refinements are most naturally applied to the 2D-tilt images, highlighting a key advantage of 2D image-based methods.

Continuous classification– machine learning-based approaches.

Machine learning architectures, particularly autoencoders⁶³, have emerged as powerful tools for analyzing cryoET data⁶⁴. In a typical autoencoder architecture, input data (i.e. 2D tilt-series images or 3D subtomograms) are passed through an encoder network, which projects the high-dimensional structural information into a simplified, low-dimensional latent space (Figure 2). This latent representation captures variations across particles, both compositional and conformational and, ideally, discards noise. A decoder network then reconstructs the input from this latent representation. Critically, this model can be trained in a self-supervised manner by comparing the decoder’s output to the input. Once trained, the autoencoder can generate particle-specific reconstructions. Notably, because each particle is mapped to a distinct point in the continuous latent space, these models avoid the rigid boundaries of discrete classification and are particularly effective in resolving continuous conformational changes and identifying sparsely populated states.

Figure 2. Autoencoder frameworks applied to cryoEM/ET datasets.

Schematic illustration of the autoencoder architecture applied to cryoEM and cryoET data, highlighting its versatility across input types (bottom). Example data types and representative tools using each are shown. The encoder (blue) projects each input into a learned low-dimensional latent space, ideally sorting individual particles according to their structural differences. Each point in the latent space corresponds to a single particle, and their distribution can reveal distinct clusters or continuous conformational landscapes. The decoder (yellow) reconstructs the input from the corresponding point in latent space, which, after dose-weighting and CTF modulation, allows for self-supervision during network training by computing the mean-squared error between the input and reconstruction. Once trained, sampling points within the latent space enables reconstruction of a continuous distribution of density maps. Adapted from²⁶.

Several recent tools adopt this autoencoder framework, though they differ in their architectural design and data handling strategies (Figure 2). For example, tomoDRGN²⁶ and cryoDRGN-ET²⁷ extend the original single-particle analysis (SPA) cryoDRGN^53,65,66 variational autoencoder framework to operate on stacks of 2D tilt-series images, allowing for users to reconstruct individual, structurally disparate particles. Note that using the 2D tilt-series data introduces specific challenges not present in SPA workflows, as each particle is associated with multiple tilt images, each with a distinct CTF, cumulative electron dose, and SNR characteristics. Additionally, some tilt images may be missing or unusable for a substantial fraction of particles analyzed due to issues associated with occlusion of the particle by another molecule along the imaging axis, poor contrast, or technical failures in image acquisition or alignment. To address this, tomoDRGN randomly samples a subset of available tilt images for each particle during training. This enforces uniform input dimensionality to the model during training while flexibly handling missing tilt-images and leveraging all available data across a tilt-series. In contrast, cryoDRGN-ET, uses a relatively small, fixed set of images - typically low-tilt angles - and discards any particles that lack this defined set of views. OPUS-TOMO⁶⁷ takes yet another approach by operating on real space 3D subtomograms using a convolutional encoder-decoder architecture (Figure 2).

To manage variation in image quality across the tilt series, tomoDRGN weights images by tilt angle to emphasize low-tilt, high-SNR views. Because the tomoDRGN model operates in reciprocal space, it also applies frequency-dependent masking to exclude high-frequency components from tilt-images degraded by high cumulative electron dose. Together, these strategies improve both reconstruction fidelity and computational efficiency²⁶, which enabled the determination of a 70S ribosome structure at 5.8Å resolution, within one week of data acquisition⁶⁸.

What lies ahead: the need for benchmarking datasets.

As seen in the adjacent fields of image processing⁶⁹ and protein structure prediction⁷⁰, the creation of robust benchmarking datasets will be critical to assess and improve tools for analyzing structural heterogeneity in cryoET. Indeed, during the development of heterogeneous structural analysis methods in SPA, public datasets^71,72 with varying degrees of structural heterogeneity served (and continue to serve) as standard benchmarks for evaluating new tools. In both instances, these datasets were especially valuable because experts deeply familiar with the systems had curated and annotated particle populations using complementary biochemical data, providing a functional “ground truth” against which newly developed tools could be easily compared. Simulated datasets^73-75 have long complemented these experimentally derived (i.e., ‘real’) ones as they naturally afford per-particle ground truth structures for assessing reconstruction accuracy, although they often fail to faithfully simulate the noise and imaging artifacts common in real data. Specific metrics, including reconstruction resolution and classification accuracy, evaluated on a per-particle basis, when ground truth is known, or in terms of the correct population frequency, have enabled comparison and iterative improvement of a large swath of heterogeneous SPA methods. Recently, such simulated and real resources were consolidated in a ‘cryoBench’⁷⁶.

A similar paradigm is emerging in cryoET, where both simulated tilt-series datasets^26,77 and increasingly well-annotated real tilt-series datasets from purified particles^8,78,79 and cells¹¹ are often analyzed. Notably, a recent large-scale cryoET study of Chlamydomonas reinhardtii cells using plasma-based focused ion beam milling produced ~1,800 tomograms, and the accompanying manuscript details expert-guided analysis of both soluble and membrane-bound complexes ranging from ~200 kDa to over 3 MDa⁵¹. As this dataset includes a variety of curated macromolecular assemblies, many of which exhibit dynamic conformational states, it is likely to be a particularly rich resource for benchmarking tomographic reconstruction and heterogeneity analysis methods.

Our group recently developed SPA benchmarking datasets that merge real and synthetic approaches: a related series of multimeric complexes differing by a single ‘programmable’ component were imaged separately and their extracted particle images were merged into unified dataset⁸⁰. Crucially, by construction, in these datasets, the identity of each particle’s variable component is known, thus providing ground truth at the level of individual particles, while still retaining the realistic noise and imaging artifacts inherent in real data. Acquiring tilt-series data of this type should be feasible, and future efforts to generate analogous datasets in cellular cryoET could serve as definitive benchmarks for evaluating tools aimed at resolving structural heterogeneity in situ. Ultimately, the continued development of such benchmarking data will be essential in motivating and evaluating forthcoming methodological innovations aimed at resolving structural dynamics in situ through cryogenic electron tomography.