Joint Angular Refinement and Reconstruction for Single-Particle Cryo-EM

Mona Zehni, Laurène Donati, Emmanuel Soubies, Zhizhen Zhao, Michael UnserIEEE TIP 2020 · OpenReview

TL;DR

This paper introduces a novel method for single-particle cryo-electron microscopy (cryo-EM) that jointly refines both the 3D density map of a molecule and the orientations of its 2D projection images. By optimizing these simultaneously on a continuous space, the approach bypasses the computationally expensive and often imprecise "template-matching" step, leading to more accurate reconstructions.

What Problem Does This Paper Address?

Cryo-EM is a powerful technique for determining the 3D structure of biological molecules. The process involves taking many 2D images of identical molecules, each at a different, unknown orientation. Reconstructing a high-resolution 3D map from these noisy 2D projections is a significant challenge. A critical step in this process is "3D refinement," where an initial, low-resolution 3D map is improved by accurately determining the orientation of each 2D image. Traditional methods often rely on "projection-matching," which involves comparing each 2D image to a large library of pre-computed templates derived from the current 3D map. This approach requires discretizing the orientation space, which can lead to inaccuracies and is computationally very expensive, especially for high-resolution reconstructions or in noisy datasets. This paper aims to overcome these limitations by developing a more efficient and accurate method for joint refinement.

Methods

The core of this work is a new optimization framework that tackles the 3D refinement problem by jointly optimizing the 3D density map and the projection orientations. The authors formulate this as a regularized least-squares minimization problem. To solve it, they employ an alternating strategy:

  • Density Map Update: The 3D density map is updated using the Alternating Direction Method of Multipliers (ADMM), a well-established algorithm for solving optimization problems by splitting them into simpler subproblems.
  • Orientation Update: The projection orientations (and in-plane translations) are refined using a semi-coordinate-wise gradient descent. A key innovation here is the derivation of explicit gradients that allow for optimization directly on the continuous orientation space, eliminating the need for discretization. This avoids the computationally intensive template-matching step.

The paper details the mathematical derivations for these gradient updates and outlines the overall algorithm, which is implemented within the GlobalBioIm library.

Datasets

The authors evaluate their method on two synthetic datasets:

  • Holliday junction complex (HJC): Used for proof-of-concept simulations and a more realistic scenario.
  • Human patched 1 (PTCH1) protein: Also used for proof-of-concept simulations.

For these datasets, they generate synthetic ground truth volumes and then simulate noisy 2D projection images under various conditions (different numbers of projections, signal-to-noise ratios, and in-plane translation magnitudes). They also use a realistic scenario with a large number of projections and a low SNR to mimic challenging experimental conditions.

Key Insights and Ideas

  • Joint Optimization is Key: The paper strongly advocates for simultaneously refining the density map and orientations, rather than treating them as separate, sequential steps. This joint approach allows information to flow bidirectionally, leading to a more accurate and robust solution.
  • Continuous Refinement: By deriving explicit gradients for orientation updates, the method operates directly on the continuous SO(3) space for rotations. This is a significant departure from traditional methods that rely on discretized grids, which can be a major source of error and computational burden.
  • Eliminating Template Matching: The proposed gradient-based approach effectively replaces the computationally expensive and potentially inaccurate template-matching step, offering a substantial speedup and improved accuracy.
  • ADMM for Density Map: Leveraging ADMM for density map updates is an efficient way to handle the reconstruction problem, especially when combined with regularization techniques.

Key Results

  • Improved Reconstruction Quality: Visual comparisons (Figure 3 and 9) show that the proposed method produces reconstructions that are significantly more accurate and closer to the ground truth than reconstructions using unrefined orientations. The quality is also shown to be comparable to, or better than, reconstructions obtained using the state-of-the-art Relion package.
  • Accurate Orientation Refinement: Probability density functions (PDFs) of orientation differences (Figure 5 and 11) demonstrate that the proposed method accurately refines projection orientations, outperforming Relion in this aspect.
  • Accurate In-Plane Translation Refinement: Similar to orientations, the method also shows strong performance in refining in-plane translations (Figure 6 and 12).
  • Competitive FSC Scores: Fourier Shell Correlation (FSC) curves (Figure 4 and 10) indicate that the proposed method achieves high resolution and its performance closely approaches that of reconstructions using perfect, ground-truth orientations.
  • Convergence: The evolution of the radial frequency at which FSC equals 0.5 (Figure 8) shows that the proposed method converges to a high-resolution map within a reasonable number of iterations, comparable to the ideal case of known orientations.

Limitations and Future Work

  • Computational Complexity: While the method avoids template matching, the overall computational complexity is still influenced by the number of projections and the size of the density map. The authors mention that the method's complexity for latent variable updates is O(PKGDn³ log n), which is competitive.
  • Hyperparameter Tuning: Like many optimization methods, the performance can be sensitive to hyperparameter choices (λ, ρ, KADMM, αθ, αt, η). The paper provides typical values used in their experiments.
  • Future Directions: The authors suggest several avenues for future work:
  • Adding regularization for latent variable estimation.
  • Combining the method with multi-scale approaches for angle refinement.
  • Further improving performance and numerical complexity.

Summary updated

Abstract

Single-particle cryo-electron microscopy (cryo-EM) reconstructs the three-dimensional (3D) structure of bio-molecules from a large set of 2D projection images with random and unknown orientations. A crucial step in the single-particle cryo-EM pipeline is 3D refinement, which resolves a high-resolution 3D structure from an initial approximate volume by refining the estimation of the orientation of each projection. In this work, we propose a new approach that refines the projection angles on the continuum. We formulate the optimization problem over the density map and the orientations jointly. The density map is updated using the efficient alternating-direction method of multipliers, while the orientations are updated through a semi-coordinate-wise gradient descent for which we provide an explicit derivation of the gradient. Our method eliminates the requirement for a fine discretization of the orientation space and does away with the classical but computationally expensive template-matching step. Numerical results demonstrate the feasibility and performance of our approach compared to several baselines.
BibTeX
@article{zehni2020joint,
  author = {Mona Zehni and Laurène Donati and Emmanuel Soubies and Zhizhen Zhao and Michael Unser},
  title = {{Joint Angular Refinement and Reconstruction for Single-Particle Cryo-EM}},
  journal = {IEEE Trans. Image Process.},
  year = {2020},
  url = {https://openreview.net/forum?id=xnX6h_gWYmE},
}
Figure