CryoDRGN

CryoDRGN employs a variational autoencoder (VAE) framework consisting of an encoder network that maps particle images to a low-dimensional latent space and a decoder network that reconstructs 3D density maps from latent coordinates. The encoder processes 2D projection images through a hierarchical pose-agnostic neural network, while the decoder uses a coordinate-based network (often a SIREN architecture with sinusoidal activations) to represent the 3D volume as a continuous function. This design enables the model to capture continuous conformational changes rather than discrete classes.

• Encoder maps images to latent embeddings z
• Decoder represents 3D density as an implicit neural representation
• Coordinate-based networks enable arbitrary resolution sampling
• Trained end-to-end with the evidence lower bound (ELBO) objective

The learned latent space encodes a continuous manifold of 3D structures, where each point corresponds to a distinct conformational state. Unlike discrete classification methods, CryoDRGN's latent space captures smooth transitions between states, revealing energy landscapes and dynamic trajectories. Users can traverse this space to visualize how domains rotate, hinges bend, or subunits rearrange. The dimensionality of the latent space (typically 8-16 dimensions) is a hyperparameter that balances expressivity against interpretability.

• Continuous manifold of conformations
• Smooth interpolation between structural states
• Dimensionality controls model capacity
• Enables discovery of rare or transient states

CryoDRGN explicitly incorporates the cryo-EM image formation model into its architecture. The decoder outputs a 3D density map, which is then projected along known viewing directions using a Fourier slice theorem implementation to generate 2D projections. These projections are compared against experimental particle images during training. The model jointly optimizes pose parameters (rotations and translations) alongside the latent conformational coordinates, using a Hartley transform for efficient, differentiable projection operations.

• Fourier-space projection for computational efficiency
• Joint optimization of pose and conformation
• Differentiable forward model enables gradient-based learning
• Accounts for contrast transfer function (CTF) corruption

CryoDRGN operates as a heterogeneous reconstruction tool that can process particle stacks after consensus refinement. The typical workflow involves: (1) extracting particles using a tool like cryoSPARC or RELION, (2) performing homogeneous refinement to obtain consensus poses, (3) training CryoDRGN on the particle stack with fixed poses, and (4) analyzing the resulting latent space. The model outputs a continuous distribution of volumes that can be sampled at any latent coordinate, enabling the generation of density maps for specific states.

• Input: particle stack with consensus poses
• Training: 10-50 epochs on GPU hardware
• Output: sampled density maps at arbitrary latent coordinates
• Compatible with .star and .cs file formats

CryoDRGN provides tools for latent space analysis through dimensionality reduction (UMAP or PCA) of the learned embeddings, revealing clusters, trajectories, and population distributions. Users can generate trajectories by interpolating between latent coordinates and visualize the resulting structural changes as movies. The k-means clustering of latent embeddings enables identification of major conformational populations, while landscape profiling quantifies the relative occupancy of each state.

• UMAP visualization of latent embeddings
• Trajectory generation via latent interpolation
• Occupancy analysis of conformational populations
• Export density maps for atomic model building

Beyond continuous domain motions, CryoDRGN can resolve compositional heterogeneity where macromolecular complexes exhibit varying subunit occupancy. The model's expressive decoder can represent both stoichiometric variation (missing or additional subunits) and continuous flexibility simultaneously. This capability is critical for studying dynamic assemblies like ribosomes, spliceosomes, and chaperone complexes where both composition and conformation vary across particles.

• Simultaneous modeling of composition and conformation
• Handles partial occupancy of subunits
• Applicable to large dynamic assemblies
• Reveals assembly and disassembly pathways

A computational classification method that sorts particle images into structurally distinct 3D classes to resolve compositional or conformational heterogeneity within a single sample. Unlike CryoDRGN's continuous latent space, traditional heterogeneous refinement produces a discrete number of 3D maps, each representing a dominant structural state. This is a prerequisite step that often motivates the use of CryoDRGN when discrete classification fails to capture subtle, continuous motions.

A per-particle, beam-induced motion correction algorithm implemented in RELION that uses a Bayesian framework to model and reverse radiation damage and movement trajectories. This preprocessing step is critical for CryoDRGN because high-resolution latent space decoding depends on accurate, high-quality particle images. Uncorrected motion blur can be misinterpreted as structural heterogeneity.

An iterative optimization algorithm used in cryo-EM refinement that alternates between computing the probability of orientation assignments (E-step) and updating the 3D density map (M-step). CryoDRGN replaces the discrete orientation search of traditional EM with a neural network that learns a continuous latent space, but the underlying probabilistic framework for image formation remains conceptually similar.