[Paper Review] Heterogeneous multireference alignment for images with application to 2-D classification in single particle reconstruction

Chao Ma, Tamir Bendory+3 more|arXiv (Cornell University)|Oct 12, 2018

Advanced Electron Microscopy Techniques and ApplicationsBiochemistry, Genetics and Molecular Biology34 references3 citations

TL;DR

This paper proposes a novel framework for 2D classification in single particle cryo-EM that directly estimates class averages from noisy, rotated particle images without clustering or rotation alignment. By leveraging rotation-invariant features like the bispectrum and solving a steerable PCA problem, the method achieves high-accuracy recovery even at low SNR, with relative estimation errors below 7% on synthetic and real molecular datasets.

ABSTRACT

Motivated by the task of 2-D classification in single particle reconstruction by cryo-electron microscopy (cryo-EM), we consider the problem of heterogeneous multireference alignment of images. In this problem, the goal is to estimate a (typically small) set of target images from a (typically large) collection of observations. Each observation is a rotated, noisy version of one of the target images. For each individual observation, neither the rotation nor which target image has been rotated are known. As the noise level in cryo-EM data is high, clustering the observations and estimating individual rotations is challenging. We propose a framework to estimate the target images directly from the observations, completely bypassing the need to cluster or register the images. The framework consists of two steps. First, we estimate rotation-invariant features of the images, such as the bispectrum. These features can be estimated to any desired accuracy, at any noise level, provided sufficiently many observations are collected. Then, we estimate the images from the invariant features. Numerical experiments on synthetic cryo-EM datasets demonstrate the effectiveness of the method. Ultimately, we outline future developments required to apply this method to experimental data.

Motivation & Objective

• Address the challenge of 2D classification in cryo-EM where particle images are noisy, randomly rotated, and unlabeled.
• Overcome limitations of existing methods like EM and MRA that rely on clustering or discrete rotation sampling, which degrade at low SNR.
• Develop a framework that estimates class averages directly from observations, bypassing the need to identify individual rotations or assign images to classes.
• Enable robust 2D classification suitable for large-scale cryo-EM datasets with high noise levels.
• Lay the foundation for future extension to continuous viewing directions, translational shifts, and real experimental artifacts like CTF and colored noise.

Proposed method

• Estimate rotation-invariant features (e.g., bispectrum) from each observed image, which are robust to unknown rotations and noise.
• Use a steerable PCA framework to represent the invariant features in a low-dimensional, rotationally invariant subspace.
• Reconstruct the original class averages by solving a non-convex optimization problem that recovers the images from the invariant features.
• Apply a one-pass algorithm over the data, making it scalable to large datasets.
• Utilize the fact that the bispectrum preserves energy and is invariant under rotation, enabling accurate feature estimation even at low SNR.
• Align recovered images to ground truth via rotation estimation after reconstruction to compute error metrics.

Experimental results

Research questions

• RQ1 Can class averages in 2D cryo-EM classification be estimated directly from noisy, rotated observations without clustering or rotation alignment?
• RQ2 How accurately can rotation-invariant features like the bispectrum be estimated from noisy data at low SNR?
• RQ3 What is the performance of a direct reconstruction framework compared to standard EM and MRA methods in terms of accuracy and computational cost?
• RQ4 How robust is the method to small translational shifts and non-uniform viewing direction distributions in real data?
• RQ5 To what extent can the method be extended to handle continuous viewing directions and experimental artifacts like CTF and colored noise?

Key findings

• The proposed method achieves a relative estimation error of 4.29% on TrpV1 data and 6.03% on yeast mitochondrial ribosome data after sPCA, with errors as low as 3.83% for high-probability classes.
• The method outperforms EM in accuracy at low SNR (1/100), especially when EM requires high discretization or many iterations.
• The algorithm is robust to small translational shifts, suggesting practical viability for real cryo-EM data.
• The framework enables direct recovery of class averages without intermediate clustering or rotation estimation, reducing error propagation.
• The method is scalable, requiring only one pass over the data, making it suitable for large-scale cryo-EM datasets.
• Theoretical and numerical results suggest the method can resolve up to 40–50 classes, consistent with known limits in 1D hMRA.