2D Class Averaging

The fundamental purpose of 2D class averaging is to overcome the extremely low signal-to-noise ratio (SNR) inherent in cryo-EM micrographs. By aligning and averaging thousands of similar particle projections, the coherent signal (the macromolecule's projection) adds constructively while the incoherent noise (shot noise, amorphous ice background) adds in quadrature. This results in a theoretical SNR improvement proportional to the square root of the number of particles averaged. A class average generated from 500 particles yields approximately a 22-fold SNR improvement over a single raw particle, making secondary structural elements like alpha-helices visually discernible for the first time.

Modern 2D class averaging employs maximum likelihood estimation (MLE) rather than simple K-means clustering. The algorithm iteratively computes the probability that each particle image belongs to each class based on cross-correlation scores, applying a regularization term (typically a Wiener filter) to prevent overfitting to noise. Key aspects include:

• Expectation-Maximization (EM): The E-step calculates class assignment probabilities; the M-step updates the class average images
• Marginalization over alignment: Each particle's translational and rotational alignment is treated as a hidden variable, integrated out probabilistically
• Resolution-dependent weighting: Higher spatial frequencies are down-weighted according to the estimated SNR at each resolution shell This probabilistic framework, implemented in RELION and cryoSPARC, is robust against the reference bias that plagued earlier deterministic approaches.

A critical secondary function of 2D classification is the computational removal of non-macromolecular contaminants. Classes that converge to featureless blobs, ice contamination, carbon edges, or aggregated denatured protein are identified and their constituent particles excluded from subsequent 3D reconstruction. This in silico purification step is essential because:

• Carbon edges produce strong, anisotropic scattering that corrupts orientation assignment
• Denatured aggregates exhibit no coherent internal features, degrading alignment accuracy
• Ice crystals introduce diffraction spots that dominate the Fourier transform
• Overlapping particles violate the single-particle assumption, producing uninterpretable averages A typical cryo-EM dataset may see 20-40% of initially picked particles discarded during this stage, dramatically improving the final map quality.

The gallery of 2D class averages provides a direct diagnostic of angular sampling completeness. Each distinct class average represents a specific projection direction of the macromolecule's 3D density. A well-behaved sample will produce class averages spanning all possible orientations, while preferred orientation artifacts manifest as a limited set of repeated views. Key indicators include:

• Top views vs. side views: Membrane proteins often adsorb to the air-water interface in a single orientation, producing only top-down projections
• Class occupancy: The number of particles assigned to each class indicates the relative frequency of that orientation
• Missing views: Gaps in the angular distribution predict anisotropic resolution in the final 3D reconstruction This diagnostic step often determines whether a dataset is suitable for high-resolution structure determination or requires sample optimization.

The translational and rotational alignment of raw particles during class averaging must achieve sub-pixel precision to avoid blurring high-resolution features. Modern implementations use Fourier-space cross-correlation with padding to achieve translational accuracy of 0.1 pixels or better. Rotational alignment employs polar Fourier transforms or adaptive band-limit schemes that progressively include higher frequencies as the class average improves. Critical algorithmic details include:

• Cubic or Lanczos interpolation: Higher-order interpolation kernels preserve high-frequency information during image rotation
• Centering heuristics: Initial translational alignment often uses a low-pass filtered reference to avoid trapping in local optima
• Symmetry handling: For particles with known point-group symmetry, rotational alignment is restricted to the asymmetric unit to avoid redundant computations Alignment errors accumulate as the square of the spatial frequency, making precise interpolation essential for achieving near-atomic resolution.

2D classification of modern cryo-EM datasets containing millions of particles is computationally intensive, requiring efficient implementations. GPU acceleration has reduced processing times from weeks to hours. Key optimization strategies include:

• Batch processing: Particles are processed in mini-batches sized to fit GPU memory (typically 500-2000 particles per batch)
• FFT-based alignment: Translational cross-correlation is computed via fast Fourier transforms, reducing complexity from O(N²) to O(N log N)
• On-the-fly downsampling: Initial iterations use heavily binned particles (e.g., 4x or 8x downsampling) to rapidly converge class averages before switching to full-resolution data
• Multi-GPU parallelization: RELION's implementation distributes particle subsets across multiple GPUs using MPI, achieving near-linear scaling A dataset of 5 million particles can be classified into 100 classes in approximately 4-6 hours on a modern 4-GPU workstation.

The computational identification and extraction of individual macromolecular projections from noisy micrographs. Modern approaches use deep learning models like Topaz and crYOLO, which are trained on annotated datasets to distinguish true particles from carbon edges and ice contamination. Accurate particle picking is the critical first step—errors here propagate through the entire pipeline, reducing the quality of 2D class averages and the final 3D reconstruction.

A mathematical function describing how the microscope's objective lens aberrations modulate image contrast as a function of spatial frequency. CTF correction must be applied to restore the true signal before 2D classification. The CTF oscillates between positive and negative contrast, creating characteristic Thon rings in the Fourier transform. Software like CTFFIND4 estimates defocus and astigmatism parameters to computationally reverse these phase flips.

The computational determination of a macromolecule's Coulomb potential density map from its 2D projections. After 2D class averaging has identified good particles and their relative orientations, algorithms like weighted back-projection or iterative refinement in RELION and cryoSPARC combine these views into a 3D volume. The process relies on the central section theorem, which states that each 2D projection corresponds to a central slice through the 3D Fourier transform.

A resolution estimation method that splits the particle dataset into two independent half-sets before 3D reconstruction. The Fourier Shell Correlation (FSC) between the two resulting maps measures the consistency of the reconstruction as a function of spatial frequency. The gold-standard procedure prevents overfitting and noise correlation, providing an unbiased resolution estimate. The 0.143 cutoff criterion is the accepted standard for reporting cryo-EM map resolution.

A computational classification method that sorts particle images into structurally distinct 3D classes. While 2D class averaging separates views and removes junk, heterogeneous refinement resolves compositional or conformational heterogeneity—such as a ligand-bound versus unbound state—that may coexist in a single sample. This is essential for studying dynamic macromolecular machines and separating functional states.

A per-particle, beam-induced motion correction algorithm implemented in RELION. It uses a Bayesian framework to model and reverse radiation damage and movement trajectories for each individual particle. By operating on the 3D reconstruction's reference projections, it achieves higher accuracy than global frame alignment alone, restoring high-resolution signal that is critical after good 2D classes have been selected.