Reconstructing continuously heterogeneous structures from single particle cryo-EM with deep generative models

Ellen D. Zhong, Tristan Bepler, Joseph H. Davis, Bonnie Berger

Introduction

Cryo-electron microscopy (cryo-EM) has revolutionized structural biology by enabling near-atomic resolution imaging of biomolecules. However, a significant challenge remains: proteins are dynamic, flexible machines, not static objects. In a typical experiment, thousands of individual protein particles are frozen in various conformational states. Traditional reconstruction methods attempt to resolve this heterogeneity by grouping images into a small number of discrete "buckets" or clusters. While effective for rigid structures, this approach fails to capture continuous motion, often resulting in blurred artifacts or a loss of structural detail for flexible regions.

The cryoDRGN Framework

To address these limitations, the authors present cryoDRGN (Deep Reconstructing Generative Networks), a deep generative model designed to reconstruct continuous distributions of 3D protein structures. Unlike traditional methods that rely on voxel-based grids, cryoDRGN represents the 3D volume as a continuous function parameterized by a neural network.

The architecture is based on a Variational Autoencoder (VAE):

  • Encoder: Maps each 2D cryo-EM image to a low-dimensional latent vector ($z$), representing its specific conformational state.
  • Decoder: A coordinate-based multi-layer perceptron (MLP) that takes a 3D coordinate in Fourier space ($k$) and the latent vector ($z$) as input to predict the corresponding intensity. This implicit representation allows for infinite resolution and enforces structural continuity across the latent space.
  • Pose Inference: To handle the unknown orientations of particles in the ice, the model employs a "Branch and Bound" global search algorithm to optimize rotation and translation parameters during training.
The cryoDRGN architecture: An encoder maps input images to a latent space $z$, while a coordinate-based decoder reconstructs the 3D volume in Fourier space using the Fourier Slice Theorem.

Results and Validation

The researchers validated cryoDRGN using both synthetic and experimental datasets. On rigid structures like the 80S ribosome, cryoDRGN achieved accuracy comparable to industry-standard tools such as cryoSPARC. More importantly, on datasets exhibiting continuous heterogeneity—such as the assembly intermediates of the E. coli large ribosomal subunit—cryoDRGN successfully mapped the full landscape of conformational states.

As shown in the comparison below, while traditional discrete clustering (cryoSPARC) identifies distinct classes, cryoDRGN provides a continuous latent representation. This allows researchers to visualize the transitions between states and identify subtle sub-populations that might be missed by manual clustering.

Comparison of reconstructions and latent space analysis. (a) cryoDRGN and cryoSPARC produce similar high-quality maps for major states. (b) The latent encoding histogram demonstrates how cryoDRGN captures the continuous distribution of particles across these states.

Discussion and Future Work

By shifting from discrete clusters to a continuous neural representation, cryoDRGN enables the discovery of complex biological motions. However, the approach is not without challenges. The global search for particle poses is computationally intensive, and achieving high-resolution details still requires substantial data and computational resources. Future developments may focus on incorporating physics-based priors or more sophisticated noise models to further enhance resolution and efficiency. Ultimately, cryoDRGN represents a major step toward viewing proteins as the dynamic ensembles they truly are.

Summary updated

Abstract

Cryo-electron microscopy (cryo-EM) is a powerful technique for determining the structure of proteins and other macromolecular complexes at near-atomic resolution. In single particle cryo-EM, the central problem is to reconstruct the three-dimensional structure of a macromolecule from $10^{4-7}$ noisy and randomly oriented two-dimensional projections. However, the imaged protein complexes may exhibit structural variability, which complicates reconstruction and is typically addressed using discrete clustering approaches that fail to capture the full range of protein dynamics. Here, we introduce a novel method for cryo-EM reconstruction that extends naturally to modeling continuous generative factors of structural heterogeneity. This method encodes structures in Fourier space using coordinate-based deep neural networks, and trains these networks from unlabeled 2D cryo-EM images by combining exact inference over image orientation with variational inference for structural heterogeneity. We demonstrate that the proposed method, termed cryoDRGN, can perform ab initio reconstruction of 3D protein complexes from simulated and real 2D cryo-EM image data. To our knowledge, cryoDRGN is the first neural network-based approach for cryo-EM reconstruction and the first end-to-end method for directly reconstructing continuous ensembles of protein structures from cryo-EM images.
BibTeX
@misc{zhong2019reconstructing,
  author = {Ellen D. Zhong and Tristan Bepler and Joseph H. Davis and Bonnie Berger},
  title = {{Reconstructing continuously heterogeneous structures from single particle cryo-EM with deep generative models}},
  year = {2019},
  url = {https://openreview.net/forum?id=SuhKkk-5GJ},
}

Loading related content...

Figure