3D Reconstruction

A foundational analytical reconstruction method that smears each 2D projection back into a 3D volume along its original projection angle. Weighting functions are applied in Fourier space to correct for the non-uniform sampling of spatial frequencies, preventing the over-representation of low-resolution components. While computationally fast, this direct inversion technique has largely been superseded by iterative methods due to its sensitivity to missing data and noise amplification in cryo-EM contexts.

The dominant statistical framework for modern 3D reconstruction, implemented in packages like RELION and cryoSPARC. The algorithm alternates between two steps:

• E-step (Expectation): Calculates the probability distribution of orientation and conformational state assignments for every particle image given the current 3D model.
• M-step (Maximization): Updates the 3D density map by a weighted back-projection of all particles, where the weights are the probabilities computed in the E-step. This maximum likelihood approach inherently handles noise and structural heterogeneity.

The definitive metric for resolution estimation that prevents overfitting. The particle dataset is randomly split into two independent halves before any 3D reconstruction begins. Two maps are reconstructed completely independently, and the Fourier Shell Correlation between them is calculated as a function of spatial frequency. The resolution is reported where the FSC curve drops below the fixed threshold of 0.143. This 'gold-standard' procedure ensures that noise is not artifactually correlated into the final map.

Computational restoration of the image signal degraded by the microscope's physics. Contrast Transfer Function (CTF) correction involves estimating the oscillating defocus and astigmatism parameters for each micrograph and computationally inverting the phase flips to restore signal. Dose weighting accounts for radiation damage by optimally down-weighting later movie frames where high-resolution information has been destroyed, implemented in motion correction software like MotionCor2 to maximize the signal-to-noise ratio of the final particle stack.

Advanced algorithms to resolve structural dynamics from a single dataset. Heterogeneous refinement in RELION performs 3D classification to sort particles into discrete conformational states. For continuous motions, 3D Variability Analysis (3DVA) in cryoSPARC uses linear principal component analysis, while deep learning tools like CryoDRGN employ variational autoencoders to learn a latent space encoding the full energy landscape of molecular motion, outputting a trajectory of 3D volumes.

Post-processing operations to enhance interpretability. Map sharpening applies a negative B-factor to the Fourier amplitudes to restore high-frequency detail dampened by the imaging process, often using a Guinier plot analysis. Local resolution estimation calculates a resolution value for every voxel in the map, generating a color-coded heatmap that identifies flexible loops or disordered domains. Tools like DeepEMhancer use convolutional neural networks to perform a non-linear, locally adaptive sharpening that mimics the appearance of atomic-resolution maps.

A direct Fourier-space reconstruction method where each 2D projection is smeared back into a 3D volume along its original projection angle. Weighting functions are applied to compensate for the non-uniform sampling of Fourier space, preventing over-representation of low-frequency components. While computationally fast, this method has been largely superseded by iterative approaches for high-resolution work due to its sensitivity to missing angular information and noise amplification.

An iterative optimization algorithm that alternates between two steps to converge on the most probable 3D structure. The E-step computes the probability distribution of orientation and class assignments for each particle image given the current 3D model. The M-step updates the 3D density map by maximizing the likelihood of observing the experimental images under those assignments. This framework, implemented in RELION, naturally handles structural heterogeneity through maximum-likelihood classification.

A rigorous resolution estimation protocol that prevents overfitting by splitting the particle stack into two independent half-sets before any 3D reconstruction begins. Two separate maps are refined completely independently, and the Fourier Shell Correlation (FSC) is computed between them. The spatial frequency at which the FSC drops below 0.143 is reported as the resolution. This 'gold-standard' procedure ensures that noise is not artificially correlated with the reconstructed signal, providing a conservative and trustworthy resolution estimate.

A computational classification strategy to resolve discrete structural variability within a single cryo-EM dataset. Particles are simultaneously aligned and sorted into a user-specified number of distinct 3D classes. This separates functionally distinct states—such as ligand-bound vs. unbound conformations or different assembly intermediates—that would otherwise average out into a blurred reconstruction. Modern implementations use maximum-likelihood or Bayesian approaches to probabilistically assign particles to classes during iterative refinement.

A method implemented in cryoSPARC that models continuous conformational landscapes rather than discrete classes. Using principal component analysis on the particle latent space, 3DVA decomposes structural heterogeneity into a linear combination of orthogonal motion components. The output is a series of eigenvolumes that describe the dominant modes of motion, enabling visualization of flexible domains, hinge motions, and continuous energy landscapes from a single cryo-EM dataset.

A deep generative model that uses a variational autoencoder (VAE) to learn a low-dimensional latent space encoding the full conformational heterogeneity present in a cryo-EM particle stack. Unlike linear methods like 3DVA, CryoDRGN can capture non-linear structural transitions and rare intermediate states. The trained decoder network can generate a 3D density map for any coordinate in the learned latent space, enabling exploration of complex, multi-state energy landscapes.