Source-linked AI summary
Rotationally Invariant Image Representation for Viewing Direction Classification in Cryo-EM
Zhizhen Zhao, Amit Singer
TL;DR
Low SNR makes it difficult to visualize individual cryo-EM particles, and global rotational alignment can be inaccurate across a wide variety of viewing angles. The paper combines steerable PCA, bispectrum features, and Vector Diffusion Maps to classify and align similar views; its procedure is reported as fast and robust to noise compared with commonly used methods.
Problem
Low SNR makes individual particles difficult to visualize, while global alignment can be inaccurate when images span a wide variety of viewing angles.
Source: §1 · p. 2, §2.1 · p. 9Method
The procedure uses Fourier-Bessel steerable PCA for image compression and denoising, bispectrum-like invariant features for classification, and Vector Diffusion Maps to improve neighbor search and rotational alignment.
Source: §1 · pp. 4–5, §5 · p. 35Results
The procedure is reported as fast and robust to noise versus commonly used class-averaging methods, and its denoising improves classification in RFA, MSA/MRA, EMAN2, and Xmipp.
Source: §5 · p. 35Takeaways & Limitations
Class averages can support common-lines-based 3D ab initio reconstruction, direct observation of heterogeneity or symmetry, and particle subgrouping for further analysis.
Source: §1 · p. 2The least-squares shift solution is exact only for images with the same viewing direction and is not expected to find shifts between different classes.
Source: §3.4 · p. 21
Abstract
from arXiv · showhide
We introduce a new rotationally invariant viewing angle classification method for identifying, among a large number of Cryo-EM projection images, similar views without prior knowledge of the molecule. Our rotationally invariant features are based on the bispectrum. Each image is denoised and compressed using steerable principal component analysis (PCA) such that rotating an image is equivalent to phase shifting the expansion coefficients. Thus we are able to extend the theory of bispectrum of 1D periodic signals to 2D images. The randomized PCA algorithm is then used to efficiently reduce the dimensionality of the bispectrum coefficients, enabling fast computation of the similarity between any pair of images. The nearest neighbors provide an initial classification of similar viewing angles. In this way, rotational alignment is only performed for images with their nearest neighbors. The initial nearest neighbor classification and alignment are further improved by a new classification method called vector diffusion maps. Our pipeline for viewing angle classification and alignment is experimentally shown to be faster and more accurate than reference-free alignment with rotationally invariant K-means clustering, MSA/MRA 2D classification, and their modern approximations.
1. Introduction
Low SNR makes individual cryo-EM particles difficult to visualize, so class averaging requires fast and accurate identification and alignment of images from similar viewing directions. The paper proposes a rotationally invariant, bispectrum-based representation to classify views before alignment, then improves the neighbor classification and alignment with Vector Diffusion Maps.
1. Introduction
- Low SNR makes individual particles difficult to visualize, so class averaging identifies, centers, aligns, and averages images from the same projection angles to improve resolution.Class averages can also support common-lines-based 3D reconstruction, heterogeneity or symmetry inspection, and particle subgrouping.Source: §1 · p. 2
- Prior approaches include MSA/MRA classification and reference-free alignment followed by rotationally invariant K-means, alongside newer methods in EMAN2, Xmipp, SPARX, and Relion.MSA/MRA compresses and denoises images before hierarchical classification and reference-based averaging; other packages use distinct invariant-feature, clustering, or Bayesian procedures.Source: §1 · pp. 2–3
- Reference-free alignment can produce large errors across many viewing directions because no assignment of in-plane rotations can align all images simultaneously.The method instead computes invariant pairwise distances, avoiding costly alignments between images from very different views.Source: §1 · p. 3
- The image representation expands images in a steerable basis and uses bispectrum features that retain phase information and specify an image up to arbitrary rotation.PCA compresses the images, and randomized low-rank approximation reduces the high-dimensional invariant feature vectors.Source: §1 · p. 4
- The proposed workflow finds rotationally invariant nearest neighbors first and computes rotational alignment angles only for those pairs.This reverses the order of alignment and classification to reduce computation; nearest-neighbor search can use a randomized algorithm for large image sets.Source: §1 · pp. 4–5
- Vector Diffusion Maps improves the initial neighbor classification by using the consistency of in-plane rotations along multiple paths through common neighbors.The introduction contrasts this with K-means clustering, whose cluster sizes can vary considerably.Source: §1 · p. 5
2. Motivation
Global reference-free alignment is inaccurate across many viewing directions, and clustering produces uneven class sizes; the section motivates classifying images through nearest neighbors instead.
2.1. No global rotational alignment
- When views cover the sphere, reference-free global alignment cannot align all images consistently because smooth tangent-plane bases would contradict the hairy ball theorem.Source: §2.1 · pp. 9–10
- 0.13, 0.09, and 0 of <5° view-pair alignments exceed 2° error for whole-sphere, 60°, and 20° caps, respectively.Source: §2.1 · p. 9, Figure 3 · p. 10
2.2. Classification instead of clustering
- K-means cluster sizes vary considerably, producing class averages with different signal-to-noise ratios and resolutions.Source: §2.2 · p. 11, Figure 4 · p. 11
- The proposed alternative averages each image with a fixed number of aligned nearest neighbors, yielding one average per image with equal SNR and preserving viewing-direction coverage.Source: §2.2 · p. 11
3. Methods
The methods build rotationally invariant image features from steerable PCA coefficients, then refine nearest-neighbor classification and alignment using path consistency and local shift estimation.
3.1. Fourier-Bessel Steerable PCA
- Fourier-Bessel steerable PCA supplies a data-adaptive basis for compressing and denoising images, with rotations represented as phase shifts of expansion coefficients.The basis is computed efficiently on Cartesian-grid images and uses rotated copies of the data to improve eigenimage estimates.Source: §3.1 · pp. 11–12
3.2. Bispectrum-like Rotationally Invariant Image Representation
- The method extends the 1D bispectrum, a shift-invariant feature that retains phase information, to rotationally invariant features for 2D images, which are expanded using a Fourier-Bessel steerable PCA basis.Source: §3.2 · pp. 13–14
- Randomized low-rank approximation reduces the high-dimensional bispectrum features, and distances between reduced representations identify nearest neighbors without in-plane alignment.The method defines image affinity by normalized cross-correlation of reduced feature vectors, then rotationally aligns images with their nearest neighbors.Source: §3.2 · p. 16
3.3. Vector Diffusion Maps Classification and Rotational Alignment
- Vector Diffusion Maps (VDM) refines neighbors by measuring agreement among in-plane rotations along multiple paths, then uses leading eigenvectors to improve rotational alignment.The affinity is larger when path transformations agree; the sparse Hermitian matrix requires O(nκ) storage.Source: §3.3 · pp. 17–20
3.4. Shift Alignment
- Because the rotationally invariant features are not shift-invariant, the method low-pass filters images and searches for shifts only within classified neighborhoods.Center-of-mass centering performs poorly at low SNR or across defocus groups; common-lines can estimate centers later.Source: §3.4 · pp. 20–21
4. Experimental Results
On simulated 70S data, the method improved viewing-angle classification over five alternatives and produced the best-matching ab initio model; experimental tests showed improved ab initio resolution for 70S and 50S and class averaging for IP3R1. Refinement outcomes varied across data sets.
4.1. Simulated noisy data
- The method's ab initio model agreed best with the simulated reference, and its classification outperformed five alternatives; generating 104 class averages took about half an hour.Refinement curves were similar across simulated-data models, though the ASPIRE starting model converged in 14 iterations and ab initio volume quality affected convergence rate.Source: §4.1 · pp. 26, 28, 29
4.2. Experimental data: 70S ribosome
- 8.58Å at the 0.143 cutoff and 10.25Å at the 0.5 cutoff were achieved by the refined 70S model; the Relion ab initio model was 15.38Å at 0.143.The two ASPIRE ab initio volumes were consistent up to 11.53Å; the prior 11.5Å result used a 0.5 cutoff but was not independently refined using gold-standard FSC.Source: §4.2 · pp. 29–30, Figure 13 · p. 31
4.3. Experimental data: 50S ribosomal subunit
- At the 0.143 cutoff, ASPIRE’s ab initio model resolved to 9.75 Å versus Xmipp’s 15.91 Å; both refined models reached 8.64 Å, taking 20 and 19 iterations, respectively.Source: §4.3 · p. 32, Figure 16 · p. 33
4.4. Experimental data: IP3R1
- The method generated class averages for IP3R1, demonstrating its application to particles with non-trivial point-group symmetry; no 3D reconstruction was attempted.The common-lines ab initio orientation procedures had not yet been modified for such symmetry.Source: §4.4 · p. 34
5. Summary and Discussion
The paper introduces a 2D class-averaging procedure combining steerable PCA, bispectrum-based rotational invariants, and Vector Diffusion Maps. Tests on simulated and experimental data report speed and noise robustness against commonly used methods, with application to particles with non-trivial point-group symmetries.
5. Summary and Discussion
- Random viewing directions make global rotational alignment impossible, so distances computed from globally aligned images are not rotationally invariant.The paper introduces a class-averaging procedure using Fourier-Bessel steerable PCA, bispectrum-like invariant features, and Vector Diffusion Maps.Source: §5 · pp. 34–35
- Fourier-Bessel steerable PCA computes eigen-images of images and their rotated copies, and offers an alternative to MSA for compression and denoising.The paper reports that this denoising improves classification results in RFA, MSA/MRA, EMAN2, and Xmipp.Source: §5 · p. 35
- Bispectrum features preserve phase information and uniquely specify an image up to rotation; dimensionality reduction and randomized nearest-neighbor search support efficient classification.The method searches alignment parameters only for nearest neighbors, reversing the usual order of alignment and classification.Source: §5 · p. 35
- For low SNR, Vector Diffusion Maps uses consistency of in-plane rotations within neighborhoods to improve initial viewing-angle classification and estimate nearest-neighbor alignments.The reported procedure is fast and robust to noise compared with methods in SPIDER, IMAGIC, EMAN2, Relion, and Xmipp, and is applicable to non-trivial point-group symmetries.Source: §5 · p. 35