Maximum Likelihood Estimation (MLE)

MLE explicitly accounts for the shot noise and detector noise inherent in low-dose cryo-EM images. Instead of treating each pixel as a perfect measurement, it models the image formation process as a probabilistic event where the observed intensity is a noisy realization of a true underlying signal. This allows the algorithm to appropriately weight high-frequency information and avoid overfitting to noise, a critical advantage over cross-correlation methods that assume noise-free references.

The true orientation, translation, and conformational state of each particle are latent (hidden) variables. MLE does not commit to a single best assignment. Instead, it computes a weighted integral over all possible assignments, where the weight is the probability that a given assignment produced the observed image. This probabilistic marginalization:

• Reduces model bias from incorrect hard assignments
• Handles structural heterogeneity naturally
• Produces smoother, more robust convergence in refinement

MLE in cryo-EM is implemented via the Expectation-Maximization (EM) algorithm, an iterative two-step process:

E-Step (Expectation): Calculate the posterior probability of every possible orientation and class assignment for each particle, given the current 3D model estimate.

M-Step (Maximization): Update the 3D reconstruction by a weighted back-projection where each particle contributes to multiple orientations proportional to its assignment probabilities.

This cycle repeats until convergence, guaranteeing that the likelihood of observing the data never decreases.

A critical implementation detail is the gold-standard refinement protocol. The particle dataset is split into two independent half-sets, and two separate 3D reconstructions are refined simultaneously. The Fourier Shell Correlation (FSC) between these independent half-maps provides an unbiased resolution estimate. Because the half-sets share no noise, any correlation above the noise floor represents true signal, preventing the algorithm from interpreting noise as structural detail—a phenomenon known as overfitting.

Pure MLE can still amplify noise at high spatial frequencies where the signal-to-noise ratio is poor. Regularized likelihood optimization (sometimes called maximum a posteriori or MAP estimation) incorporates prior knowledge about the expected power spectrum of the macromolecule. This Bayesian prior acts as a smoothness constraint, suppressing noise amplification while preserving genuine structural features. The regularization parameter, often expressed as a B-factor or weight on the prior, is typically estimated automatically from the data.

RELION (REgularized LIkelihood OptimizatioN) is the seminal software package that popularized the Bayesian approach to MLE in cryo-EM. Its key innovations include:

• An empirical Bayesian prior derived from the average power spectrum of the reconstructed map
• Adaptive regularization that varies with resolution
• Integrated 3D classification to separate heterogeneous populations
• A statistical movie processing framework for per-particle motion correction (Bayesian polishing)

RELION's approach has become the standard against which other refinement packages are measured.

A multi-class extension of MLE that simultaneously solves for both orientation and structural class assignments. Rather than assuming all particles represent a single 3D structure, the likelihood function includes a discrete class variable, allowing the algorithm to sort particles into distinct conformations.

• Resolves compositional heterogeneity (e.g., bound vs. unbound ligand)
• Each class gets its own 3D reconstruction
• Competing models explain the data; MLE selects the most probable assignment per particle