Use CryoDRGN Online
Commercially Available CryoDRGN No-Code Web Server
CryoDRGN: Deep Neural Network Reconstruction of Heterogeneous Cryo-EM Structures
CryoDRGN (Deep Reconstructing Generative Networks) is a deep generative method that reconstructs continuous distributions of 3D density maps from heterogeneous single-particle cryo-EM datasets and maps per-particle heterogeneity. It was developed by Ellen D. Zhong, Tristan Bepler, Bonnie Berger and Joseph H. Davis at the Massachusetts Institute of Technology and published in Nature Methods in 2021. The authors describe it as open-source software, freely available at cryodrgn.csail.mit.edu.
Proteins and their complexes are dynamic machines, yet a single static structure hides the range of conformations relevant to function. Single-particle cryo-EM captures millions of snapshots, each of a unique molecule in its own state, so it holds promise for revealing conformational landscapes, but reconstructing ensembles of 3D volumes from those snapshots remains a major computational challenge.
Existing tools make limiting assumptions: 3D classification assumes a mixture of discrete classes, multi-body refinement assumes motions composed of user-defined rigid bodies, and PCA-based approaches such as 3D Variability Analysis assume linear combinations of density maps, which can create artifacts when motions are poorly approximated by linear interpolation. CryoDRGN instead uses neural networks to model complex, nonlinear heterogeneity without such simplifying assumptions.
Property | Detail |
|---|---|
Architecture | Image encoder and volume decoder, trained as a variational autoencoder in the Fourier domain |
Inputs | 2D particle images plus orientations (poses) from a consensus reconstruction |
Outputs | Per-particle latent encodings and a decoder that generates a 3D density map for any latent value |
User-defined setting | Dimensionality of the latent space |
Tools | Latent-space visualization, density map generation, particle subset extraction, trajectory generation |
How CryoDRGN Learns Cryo-EM Conformational Landscapes
Encode: The encoder maps each 2D particle image to a point in a continuous latent space.
Decode: The decoder generates a 3D density map from any latent value, with images treated as planar slices of the 3D volume in Fourier space at poses fixed from a consensus reconstruction.
Train: Both networks are trained jointly from random initialization with stochastic gradient descent, maximizing a variational bound on data likelihood as in standard variational autoencoders.
Explore: Representative structures are generated from dense regions of latent space (for example k-means cluster centers), and trajectories are built by sampling along paths through latent space.
Validate and filter: Any generated volume can be checked by traditional reconstruction using nearby particles, and regions enriched in impurities or artifacts can be removed from later analysis.
Results: Hidden States in Ribosome, RAG and Spliceosome Datasets
The authors first showed that the decoder can represent single high-resolution maps, matching traditional reconstructions up to 3.6 Angstrom for the RAG1-RAG2 complex and 3.9 Angstrom for the Plasmodium falciparum 80S ribosome (FSC = 0.5). They then applied the full method to four public datasets and to simulated datasets with uniform, cooperative, noncontiguous and compositional heterogeneity.
Dataset | Finding |
|---|---|
Pf80S ribosome (EMPIAR-10028) | A rotated 40S state was found; 4,889 particles from the outlying cluster were validated by cryoSPARC refinement |
RAG1-RAG2 (EMPIAR-10049) | Residual heterogeneity revealed in a dataset treated as homogeneous |
Assembling E. coli 50S subunit (EMPIAR-10076) | A previously unseen state was found, supported by 1,113 selected particles |
Pre-catalytic spliceosome (EMPIAR-10180) | Large continuous motion of the SF3b subcomplex and helicase reconstructed, without the rigid-body boundary artifacts of multi-body refinement |
The paper reports that rare states containing as few as about 1,000 particles can be identified and that impurities can be filtered using the latent space.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with CryoDRGN on Tamarind Bio
Continuous conformational motion: Visualize large-scale motions of complexes such as the spliceosome as trajectories rather than a few discrete classes.
Rare and hidden states: Uncover residual heterogeneity in datasets that appear homogeneous.
Assembly intermediates: Model compositional changes in assembling complexes without specifying the number of states in advance.
Dataset cleanup: Identify and filter impurities or artifacts by selecting regions of latent space.
How to Use CryoDRGN on Tamarind Bio
Open the tool: Log in at tamarind.bio and open CryoDRGN.
Provide particle images: Upload the particle dataset as the tool requests.
Provide poses: Supply orientations from a consensus reconstruction, as the method requires, if exposed.
Set parameters: Choose settings such as latent dimensionality, network size and number of epochs, if exposed.
Run: Launch training and wait for the job to finish.
Download outputs: Retrieve latent encodings, generated density maps and trajectories.
Interpret: Inspect the distribution of latent encodings and k-means cluster volumes to identify states.
Use downstream: Extract particle subsets and confirm states with traditional reconstruction in your usual cryo-EM software.
Things to Keep in Mind
The objective only reproduces the distribution of structures, so the latent space layout is not guaranteed to reflect the true energy landscape.
Structures generated from unoccupied regions of latent space generally do not correspond to real physical structures.
Poses are fixed from a consensus reconstruction, so results depend on their quality.
Larger networks represent higher resolution but trade off against training speed.
Source: Zhong ED, Bepler T, Berger B, Davis JH. CryoDRGN: Reconstruction of heterogeneous cryo-EM structures using neural networks. Nature Methods 2021; 18(2):176-185. DOI: 10.1038/s41592-020-01049-4. Project site: cryodrgn.csail.mit.edu.