CryoDRGN

CryoDRGN: reconstruction of heterogeneous cryo-EM structures using neural networks. Zhong ED, Bepler T, Berger B, Davis JH. Nat Methods. 2021 Feb;18(2):176-185.

  • Cryo-EM single-particle analysis has proven powerful in determining the structures of rigid macromolecules. However, many complexes exhibit conformational and compositional heterogeneity that pose a major challenge to existing 3D reconstruction methods.

  • cryoDRGN uses deep neural networks to reconstruct continuous distributions of 3D density maps reflecting the heterogeneity within single-particle cryo-EM datasets.

  • CryoDRGN is open-source software freely available from http://cryodrgn.csail.mit.edu

[back to paper list]

Existing Methods: Homogeneous vs. Heterogeneous


  • Homogeneous reconstruction considers the data to represent a single conformational (and compositional) state
  • Many methods require classifying particles into quasi-homogeneous subsets, then using the different classes for separate reconstructions


Problems with this approach:

  • The number of states/classes is not known a priori and may be difficult to determine accurately
  • Using a discrete number of classes does not account for continuous variations in conformation

More-Advanced Heterogeneous Methods


  • RELION multibody refinement – models the structure as the sum of user-defined rigid bodies that can rotate relative to one another

  • cryoSPARC 3D variability analysis – PCA-based, represents variability within a linear subspace (which may be misleading when structural deformations are not well described by linear interpolations between basis vectors)

  • manifold embedding – binning particles along the data manifold followed by traditional homogeneous reconstruction

  • others...

CryoDRGN (Deep Reconstructing Generative Networks) Method


  • unsupervised learning by two neural networks, encoder and decoder

  • encodes 2D images into a latent space from which 3D maps are reconstructed

  • assumes structures fall along a low-dimensional manifold in the latent space
    • dimensionality specified by the user, 10 dimensions generally effective

  • density maps can be generated from arbitrary values of the latent variable

  • regions of latent space rich in impurities or artifacts can be filtered out
    • initial filtering run can be done on downsampled data (Fourier clipping)

What does CryoDRGN allow researchers to do?


  • get latent-space encoding of particle distribution (which can be embedded in 2D for visualization, e.g. with PCA or UMAP)

  • generate density maps for exploratory analysis

  • extract particle subsets for use with other tools

  • generate trajectories of molecular motions (as maps)

  • mainly, to characterize cryoEM dataset heterogeneity and generate maps of the various states without having to classify the particles ahead of time

“Homogeneous” High-Resolution Data

EMPIAR (Electron Microscopy Public Image ARchive) – raw data for some EMDB entries. 2022-08-02: 975 datasets, 2.0 PB. For RAG1-RAG2 (EMPIAR 10049), P. falciparum 80S ribosome (EMPIAR 10028):

  • A. cryoDRGN (1024 x 3, "no latent variable input") vs. traditional (cryoSPARC)
  • B. FSC (Fourier shell correlation) curves for NN architectures (E), resln <4Å at FSC 0.5
  • C. FSC evolution over training epochs, D. similar for loss
  • E. Architecture is tradeoff between resolution and speed
  • F. CryoDRGN map with RAG1-RAG2 atomic structure (PDB 3jbx)

Simulated Heterogeneity

  • A. Conformational: Uniform sampling along 1D reaction coordinate of rotating a single dihedral, Cooperative biased to specific states, Noncontiguous with unobserved transition states
  • B. Compositional from mixing bacterial 30S, 50S, and 70S ribosomal complexes
  • C. for the Uniform data, sample of cryoDRGN maps and per-image FSCs (output maps vs. ground-truth maps across the reaction coordinate)
  • D. CryoDRGN maps from the Compositional data
  • E-H. Latent-space encodings (using 1D model) consistent with ground truth

Heterogeneity in “Homogeneous” Data

  • A. RAG1-RAG2 maps from original authors based on 2 particle classes: signal end (only core resolved) and paired (other parts visible)
  • B,C. In the signal end data, cryoDGRN (10D) finds positions of formerly unresolved parts (ribbons show the previous paired structure for comparison) possibly important for function [video]
  • D,E. In the Pf80S ribosome data, cryoDRGN (10D) identifies a rotated-subunit substate

Highly Heterogeneous Data

  • – E. coli large ribosomal subunit (50S) undergoing assembly (EMPIAR-10076)
  • – Original analysis found 13 classes grouped into 4 major assembly states
  • A-D. CryoDRGN 1D and 10D latent space representations and filtering (details in suppl fig 5)
  • E-G. CryoDRGN finds the 4 states without prior classification (dashed outline shows mature subunit)
  • F,G. Latent space UMAP colored/labeled by states from original analysis
  • H. Some particles ("A" in panel G) are whole ribosome (70S = 30S + 50S)
  • I. C4 is previously unreported assembly intermediate

Large Continuous Conformational Changes

  • – Pre-catalytic spliceosome (EMPIAR-10180)
  • – Original expert classification suggested a continuum of conformations with large motions of the SF3b subcomplex
  • A,B. CryoDRGN (10D) finds several clusters; filter to remove those with likely imaging artifacts (iii), missing the SF3b subcomplex (iv), or that have the additional U2 core (v)
  • C,D. Maps generated along principal components show large movements of the SF3B subcomplex relative to the rest of the structure [video]

Others Can Use CryoDRGN

Can We Use CryoDRGN in ChimeraX?

  • ...or its outputs after training?

  • related ticket #6106: Efficiently select maps by visual inspection