Recovering a Molecule's 3D Dynamics from Liquid-phase Electron Microscopy Movies

Enze Ye, Yuhang Wang, Hong Zhang, Yiqin Gao, Huan Wang, He SunICCV 2023 · DOI · OpenReview

Abstract

The dynamics of biomolecules are crucial for our understanding of their functioning in living systems. However, current 3D imaging techniques, such as cryogenic electron microscopy (cryo-EM), require freezing the sample, which limits the observation of their conformational changes in real time. The innovative liquid-phase electron microscopy (liquid-phase EM) technique allows molecules to be placed in the native liquid environment, providing a unique opportunity to observe their dynamics. In this paper, we propose TEMPOR, a Temporal Electron MicroscoPy Object Reconstruction algorithm for liquid-phase EM that leverages an implicit neural representation (INR) and a dynamical variational auto-encoder (DVAE) to recover time series of molecular structures. We demonstrate its advantages in recovering different motion dynamics from two simulated datasets, 7bcq & Cas9. To our knowledge, our work is the first attempt to directly recover 3D structures of a temporally-varying particle from liquid-phase EM movies. It provides a promising new approach for studying molecules’ 3D dynamics in structural biology.

Introduction

In structural biology, Cryogenic Electron Microscopy (cryo-EM) is the gold standard for visualizing molecules at atomic resolution. However, the requirement to freeze samples essentially "stops the clock," preventing researchers from observing how proteins move and interact in their native states. Liquid-phase EM (LP-EM) offers a solution by maintaining molecules in a liquid environment, but the resulting data is extremely noisy and sparse. Unlike cryo-EM, which uses millions of static images, LP-EM provides only a few thousand frames of a moving, changing molecule. Traditional 3D reconstruction methods fail under these conditions because they cannot handle a "moving target" with only one viewing angle per time step.

Overview of TEMPOR: Reconstructing 3D molecular volumes from noisy liquid-phase EM movie frames.

Methods

We introduce TEMPOR (Temporal Electron MicroscoPy Object Reconstruction), which treats 3D reconstruction as a dynamic tomographic problem. The architecture is built upon two primary technical pillars:

  1. Implicit Neural Representation (INR): Rather than using a traditional voxel grid, TEMPOR employs a coordinate-based neural network to model the molecule's 3D scattering potential in the Fourier domain. This approach acts as a natural regularizer, providing robustness against the extreme noise levels typical of LP-EM.
  2. Dynamical Variational Auto-Encoder (DVAE): To address the temporal changes in molecular structure, the model utilizes a bidirectional Recurrent Neural Network (RNN) as an encoder. This allows the system to process the entire movie sequence, leveraging temporal continuity to infer the structural state of a molecule at any given frame by "borrowing" information from adjacent time steps.
The TEMPOR architecture, featuring an implicit neural representation (INR) generator and a dynamical variational auto-encoder (DVAE) with a bidirectional RNN.

The training process involves a pretrain-finetune strategy, where a baseline static network is established before enabling temporal dynamics, ensuring the optimization process avoids local minima.

Experimental Evaluation

The method was evaluated using two simulated datasets designed to replicate the challenges of experimental LP-EM:

  • 7bcq: A small protein undergoing simple, one-dimensional rigid-body rotation (9° per frame).
  • Cas9: A complex protein undergoing non-rigid deformational motion, transitioning from a "free" state to an "RNA-bound" state across 500 synthetic movies of 19 frames each.

These simulations provided a "ground truth" necessary for calculating resolution and accuracy, metrics that are currently difficult to verify with experimental data.

Results and Discussion

TEMPOR demonstrated significant improvements over existing methods like CryoDRGN. On the complex Cas9 dataset, the algorithm improved reconstruction resolution by an average of 13.5%. Qualitatively, TEMPOR recovered fine structural details, such as gaps between protein domains, that were previously lost to blurring.

The model proved remarkably robust to noise, maintaining stability at a Signal-to-Noise Ratio (SNR) of 0.1. Furthermore, latent space analysis revealed that TEMPOR produces more compact and clearly clustered representations of conformational states, suggesting a superior understanding of the underlying physical motion compared to static-frame approaches.

Conclusion and Future Work

TEMPOR represents the first successful attempt to directly recover 3D structures of temporally-varying particles from LP-EM movies. While the current study relies on simulated data, it paves the way for real-world applications. Future work will focus on addressing experimental hurdles such as beam-induced motion and sample damage, as well as optimizing the computational cost of training these complex neural representations. This work provides a vital new tool for studying the 3D dynamics of molecules in their native liquid environments.

Summary updated

Abstract

The dynamics of biomolecules are crucial for our understanding of their functioning in living systems. However, current 3D imaging techniques, such as cryogenic electron microscopy (cryo-EM), require freezing the sample, which limits the observation of their conformational changes in real time. The innovative liquid-phase electron microscopy (liquid-phase EM) technique allows molecules to be placed in the native liquid environment, providing a unique opportunity to observe their dynamics. In this paper, we propose TEMPOR, a Temporal Electron MicroscoPy Object Reconstruction algorithm for liquid-phase EM that leverages an implicit neural representation (INR) and a dynamical variational auto-encoder (DVAE) to recover time series of molecular structures. We demonstrate its advantages in recovering different motion dynamics from two simulated datasets, 7bcq & Cas9. To our knowledge, our work is the first attempt to directly recover 3D structures of a temporally-varying particle from liquid-phase EM movies. It provides a promising new approach for studying molecules’ 3D dynamics in structural biology.
BibTeX
@misc{ye2023recovering,
  author = {Enze Ye and Yuhang Wang and Hong Zhang and Yiqin Gao and Huan Wang and He Sun},
  title = {{Recovering a Molecule's 3D Dynamics from Liquid-phase Electron Microscopy Movies}},
  year = {2023},
  eprint = {2308.11927},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2308.11927},
}
Figure