Subtomogram Averaging

STA resolves structures directly within the cellular context, bypassing the need for biochemical purification. Unlike single-particle analysis (SPA), which requires purified, homogeneous samples, STA targets macromolecules embedded in pleomorphic environments like organelles, membranes, or viral factories.

• Preserves native protein-lipid and protein-protein interactions
• Resolves structures in functional states impossible to purify
• Enables structural biology of transient complexes and low-abundance targets

The STA workflow begins with tilt-series acquisition and tomogram reconstruction. Particles are identified via template matching or manual picking, then extracted as 3D sub-volumes. These sub-volumes undergo iterative alignment and classification to resolve structural heterogeneity.

• Template matching: Cross-correlation of a reference against the tomogram
• Missing wedge compensation: Corrects for the wedge-shaped gap in Fourier space inherent to limited tilt angles
• Dose weighting: Optimally down-weights later tilt images to account for cumulative radiation damage

Alignment in STA must account for the missing wedge of Fourier space information—a systematic artifact caused by the limited tilt range (typically ±60°) during tomogram acquisition. This wedge distorts sub-volumes in a direction-dependent manner.

• Constrained cross-correlation: Restricts alignment comparisons to mutually sampled Fourier regions
• Wedge-masked refinement: Applies a per-particle mask in Fourier space to prevent noise correlation
• RELION-4.0 and M implement wedge-aware maximum likelihood approaches for robust alignment

Cellular samples exhibit compositional and conformational heterogeneity—different binding partners, ligand states, or flexible domains coexist in the same tomogram. STA uses 3D classification to computationally separate these states.

• Maximum likelihood classification: Assigns sub-volumes to distinct structural classes based on probability
• Multi-body refinement: Models continuous motion of rigid domains using focused masks
• Principal component analysis: Captures continuous conformational landscapes, analogous to 3DVA in SPA

Resolution in STA is assessed using the gold-standard Fourier Shell Correlation (FSC) criterion. The dataset is split into two independent half-sets, reconstructed separately, and compared in Fourier space to avoid overfitting.

• FSC=0.143: The standard threshold for reporting resolution
• Local resolution estimation: Maps spatial variation in resolution, identifying flexible regions
• Tilt validation: Checks for overfitting by comparing reconstructions from odd/even tilt images

Several specialized software packages implement STA pipelines, each with distinct algorithmic approaches. The field has converged on RELION and M as primary tools, with cryoSPARC and Dynamo offering complementary capabilities.

• RELION-4.0: Implements wedge-aware Bayesian refinement with multi-body analysis
• M: Uses a modular, Python-based framework for flexible STA workflows
• Dynamo: Provides a MATLAB-based environment with extensive visualization tools
• emClarity: Specializes in high-resolution STA with CTF correction at the sub-volume level

A fundamental computational challenge in STA caused by the limited tilt range (typically ±60°) during tomogram acquisition. This results in a wedge-shaped region of missing Fourier space information, leading to anisotropic resolution and elongation artifacts in the reconstructed density. Correction methods include:

• Constrained cross-correlation during alignment that restricts comparisons to mutually sampled regions
• Compensating for the missing wedge in 3D reconstruction by weighting Fourier components based on their sampling frequency
• CTF correction applied per-tilt to restore information transfer

The initial step in STA that locates macromolecular complexes within the crowded, noisy environment of a tomogram. A reference template—often a low-pass filtered atomic model or a previously determined structure—is cross-correlated against the tomogram volume. Peaks in the correlation map indicate candidate particle positions. Modern approaches use deep learning-based picking (e.g., DeepFinder, crYOLO adapted for tomograms) to improve sensitivity and reduce false positives in the presence of cellular clutter.

The conceptual precursor to STA. In SPA, purified macromolecules are imaged as isolated 2D projections and computationally combined into a 3D reconstruction. STA extends this principle to in situ structural biology by treating extracted subvolumes as 'single particles.' Key differences:

• SPA particles are 2D projections; STA particles are 3D sub-tomograms
• SPA assumes identical copies; STA must handle conformational heterogeneity within the cellular context
• STA must correct for the missing wedge and lower signal-to-noise ratio inherent to tomograms

The statistical framework underpinning modern STA refinement, as implemented in RELION-4.0. MLE iteratively determines the set of particle orientations, translations, and class assignments that maximize the probability of observing the experimental subvolumes given a 3D reference. This approach:

• Marginalizes over hidden variables (orientation, class) rather than assigning hard classifications
• Weights particles by their information content, naturally down-weighting noisy or damaged subvolumes
• Integrates CTF and missing wedge compensation directly into the probabilistic model

A method implemented in cryoSPARC that resolves continuous conformational heterogeneity within a subtomogram dataset. Using principal component analysis in the space of 3D density, 3DVA learns a low-dimensional latent representation of structural variability. The output is a series of eigenvolumes that describe the principal modes of motion, enabling the visualization of continuous molecular motions—such as ribosome ratcheting or transporter conformational cycles—directly from in situ data.