CryoGAN: A New Reconstruction Paradigm for Single-Particle Cryo-EM via Deep Adversarial Learning

Harshit Gupta, Michael T. McCann, Laurène Donati, Michaël UnserbioRxiv 2020 · DOI

TL;DR

CryoGAN introduces a novel, unsupervised approach to reconstructing 3D molecular structures from cryo-electron microscopy (cryo-EM) images. By leveraging Generative Adversarial Networks (GANs), it bypasses the computationally intensive pose estimation step, offering faster and more accessible 3D structure determination.

What Problem Does This Paper Address?

Single-particle cryo-electron microscopy (cryo-EM) is a powerful technique for determining the atomic structure of biological molecules. However, a major hurdle is that the orientation (pose) of each individual molecule in the captured 2D images is unknown. Traditional methods tackle this by using a "maximum-likelihood" approach, which involves calculating probabilities across all possible orientations for every image. This is computationally very demanding and time-consuming. The authors aim to overcome this by developing a method that learns the 3D structure directly from the statistical properties of the images, without explicitly determining each particle's orientation.

Methods

The core innovation is CryoGAN, which adapts the Generative Adversarial Network (GAN) framework. In a standard GAN, a "generator" creates synthetic data, and a "discriminator" tries to distinguish it from real data. CryoGAN replaces the generator with a "cryo-EM physics simulator." This simulator takes a candidate 3D structure and generates simulated 2D projections that mimic the real cryo-EM imaging process (including noise and contrast transfer function effects).

The GAN then trains by:

  1. Discriminator: Learning to distinguish between real cryo-EM projection images and simulated projections from the current 3D structure estimate.
  2. Simulator (Generator): Learning to produce simulated projections that are as close as possible to the real data's distribution, thereby improving the 3D structure estimate.

This adversarial process allows CryoGAN to learn the 3D structure whose simulated projections best match the statistical distribution of the real experimental data, effectively bypassing the need to explicitly estimate individual particle poses. The method is unsupervised, requiring only projection images and CTF parameters, and does not need an initial 3D volume estimate.

Datasets

  • Synthetic B-galactosidase dataset: Generated in silico to mimic real cryo-EM data, including noise and CTF effects. This dataset was used to test the algorithm's performance and stability.
  • Real B-galactosidase dataset (EMPIAR-10061): A publicly available dataset used to demonstrate CryoGAN's ability to reconstruct structures from experimental data.

Key Insights and Ideas

  • GANs for structural biology: The paper demonstrates that GANs, primarily known for image generation, can be effectively repurposed for complex scientific reconstruction tasks like cryo-EM.
  • Distributional matching as a paradigm shift: Instead of focusing on individual particle poses, CryoGAN matches the overall statistical distribution of simulated projections to real data. This is a fundamental change from traditional likelihood-based methods.
  • Likelihood-free reconstruction: By avoiding explicit likelihood calculations and pose estimation, CryoGAN significantly reduces computational complexity and user intervention.
  • Unsupervised learning for structure determination: The ability to reconstruct a 3D structure without prior knowledge of the molecule's shape is a significant advantage, especially for novel or poorly characterized samples.

Key Results

  • Synthetic data: CryoGAN achieved a resolution of 8.6 Å on a realistic synthetic B-galactosidase dataset.
  • Real data: On the EMPIAR-10061 B-galactosidase dataset, CryoGAN achieved a resolution of 12.1 Å in approximately 150 minutes on a high-end GPU.
  • Speed: Reconstructions can be obtained in a matter of hours on a high-end GPU, a significant speedup compared to traditional methods.
  • Mathematical guarantees: The authors provide theoretical guarantees for the recovery of the correct structure under certain assumptions.

Limitations and Future Work

  • Resolution: While promising, the current resolution achieved on real data (12.1 Å) is not yet competitive with state-of-the-art methods, which can reach sub-Ångström resolutions.
  • Noise and aberrations: The method's performance can be affected by very low SNR or significant image aberrations.
  • Future directions:
    • Coarse-to-fine strategy: Implementing multi-scale GAN approaches for progressive reconstruction.
  • Improved physics simulator: Incorporating more realistic noise models (e.g., Poisson noise) and potentially simulating raw micrographs instead of individual projections.
  • Handling conformational heterogeneity: Extending the framework to reconstruct structures with multiple conformational states.
  • Learning pose distributions: Investigating methods to learn pose distributions directly, rather than assuming them.
  • Outlier detection: Developing mechanisms to handle datasets with a high proportion of corrupted particle images.

Summary updated

Abstract

We present CryoGAN, a new paradigm for single-particle cryo-EM reconstruction based on unsupervised deep adversarial learning. The major challenge in single-particle cryo-EM is that the imaged particles have unknown poses. Current reconstruction techniques are based on a marginalized maximum-likelihood formulation that requires calculations over the set of all possible poses for each projection image, a computationally demanding procedure. CryoGAN sidesteps this problem by using a generative adversarial network (GAN) to learn the 3D structure that has simulated projections that most closely match the real data in a distributional sense. The architecture of CryoGAN resembles that of standard GAN, with the twist that the generator network is replaced by a model of the cryo-EM image acquisition process. CryoGAN is an unsupervised algorithm that only demands projection images and an estimate of the contrast transfer function parameters. No initial volume estimate or prior training is needed. Moreover, CryoGAN requires minimal user interaction and can provide reconstructions in a matter of hours on a high-end GPU. In addition, we provide sound mathematical guarantees on the recovery of the correct structure. CryoGAN currently achieves a 8.6 Å resolution on a realistic synthetic dataset. Preliminary results on real β -galactosidase data demonstrate CryoGAN’s ability to exploit data statistics under standard experimental imaging conditions. We believe that this paradigm opens the door to a family of novel likelihood-free algorithms for cryo-EM reconstruction.
BibTeX
@misc{gupta2020cryogan,
  author = {Harshit Gupta and Michael T. McCann and Laurène Donati and Michaël Unser},
  title = {{CryoGAN: A New Reconstruction Paradigm for Single-Particle Cryo-EM <i>via</i> Deep Adversarial Learning}},
  year = {2020},
  doi = {10.1101/2020.03.20.001016},
  howpublished = {bioRxiv},
}

Loading related content...

Figure