Expectation-Maximization (EM)

The Expectation step calculates the probability that each 2D particle image corresponds to a specific 3D orientation and conformation. Instead of making a single hard assignment, it computes a weighted posterior distribution over all possible projection angles using the current 3D map estimate.

• Integrates over rotational (SO(3)) and translational degrees of freedom
• Uses maximum-likelihood to handle structural heterogeneity
• Outputs a responsibility matrix linking every particle to every orientation

The Maximization step updates the 3D Coulomb potential density map using the probabilistic weights from the E-step. All 2D particle images are back-projected into a 3D volume, weighted by their assignment probabilities.

• Employs Fourier-space reconstruction for computational efficiency
• Incorporates CTF correction to deconvolve lens aberrations
• Produces an updated 3D map that better explains the experimental data

Modern implementations like RELION (REgularized LIkelihood OptimizatioN) extend EM with Bayesian priors on the Fourier components of the 3D map. This prevents overfitting to noise by penalizing high-frequency components that lack statistical support.

• Assumes a Gaussian prior on signal power in each frequency shell
• Automatically adapts regularization weight using empirical Bayes
• Eliminates the need for ad-hoc low-pass filtering during refinement

To avoid overfitting and noise correlation, the dataset is split into two independent half-sets processed separately. The Fourier Shell Correlation (FSC) between the two resulting half-maps provides an unbiased resolution estimate.

• The FSC=0.143 criterion defines the reported resolution
• Prevents the algorithm from fitting noise as signal
• Serves as the objective function for monitoring EM convergence

Standard EM assumes a discrete set of rigid 3D classes. Extensions like 3D Variability Analysis (3DVA) and cryoDRGN replace the discrete M-step with a continuous generative model, learning a latent space of conformational motions.

• Uses variational autoencoders to parameterize structural flexibility
• Enables visualization of continuous domain motions
• Resolves dynamic complexes without imposing discrete state boundaries

EM refinement is computationally intensive, requiring GPU acceleration for practical execution. The algorithm iterates between E and M steps until the FSC curve converges.

• RELION: CPU/GPU hybrid with Bayesian regularization
• cryoSPARC: GPU-native with stochastic gradient descent variants
• Typical refinements require 25-50 iterations over 10⁵–10⁶ particles
• Memory footprint scales with particle count and box size

The statistical framework that EM algorithms operationalize. MLE finds the 3D structural model and orientation parameters that maximize the probability of observing the experimental particle images.

• Iteratively refines both the map and assignment probabilities
• Handles noisy, low-contrast images where deterministic alignment fails
• Forms the theoretical backbone of RELION's 3D auto-refine
• Produces resolution estimates that avoid overfitting

The resolution estimation method that prevents overfitting during EM refinement. The particle dataset is split into two independent half-sets, each refined separately, and their Fourier shells are compared.

• Gold-standard FSC = 0.143 criterion defines nominal resolution
• Ensures the EM algorithm hasn't fit noise rather than signal
• Requires completely independent half-set refinements from the start
• Standard reporting requirement for all cryo-EM structures

Extends EM to classify particles into multiple structurally distinct 3D classes simultaneously. Essential for resolving compositional or conformational heterogeneity within a single sample.

• Uses a multi-reference EM approach with competitive assignment
• Separates particles by structural state, not just orientation
• Critical for dynamic complexes with multiple conformations
• Implemented in both RELION (3D classification) and cryoSPARC

A cryoSPARC method that models continuous conformational landscapes rather than discrete classes. Uses principal component analysis within a linear subspace model to capture the full spectrum of structural motions.

• Outputs a motion trajectory along principal components
• Complements discrete EM classification with continuous dynamics
• Reveals functional motions like domain rotations and hinge movements
• Generates interpolated 3D movies of conformational changes

A per-particle beam-induced motion correction algorithm in RELION that uses a Bayesian framework to model and reverse radiation damage trajectories. Operates after initial EM refinement to further improve resolution.

• Models each particle's individual motion track through the ice
• Applies dose-dependent B-factor weighting to each movie frame
• Reverses radiation damage effects on high-resolution features
• Typically yields 0.2-0.5 Å resolution improvement

A deep generative model using a variational autoencoder to reconstruct continuous conformational heterogeneity. Unlike traditional EM, it learns a latent space of structural states directly from particle images.

• Encodes each particle image into a low-dimensional latent coordinate
• Decodes latent coordinates into 3D density maps via a neural network
• Captures complex, non-linear conformational landscapes
• Complements EM-based refinement with deep learning flexibility