Emerging Themes in CryoEM─Single Particle Analysis Image Processing

Abstract

Cryo-electron microscopy (CryoEM) has become a vital technique in structural biology. It is an interdisciplinary field that takes advantage of advances in biochemistry, physics, and image processing, among other disciplines. Innovations in these three basic pillars have contributed to the boosting of CryoEM in the past decade. This work reviews the main contributions in image processing to the current reconstruction workflow of single particle analysis (SPA) by CryoEM. Our review emphasizes the time evolution of the algorithms across the different steps of the workflow differentiating between two groups of approaches: analytical methods and deep learning algorithms. We present an analysis of the current state of the art. Finally, we discuss the emerging problems and challenges still to be addressed in the evolution of CryoEM image processing methods in SPA.

Figure 1

Yearly evolution of (left axis) the highest resolution achieved by CryoEM and (right axis) the number of deposited structures in the PDB by experimental method. Data extracted from ref (16).

Figure 2

Scheme of a basic neuron: (left) Mathematically, it is composed by multiple inputs x_i, which are linearly combined applying different weights, w. This linear combination is the argument of the activation function g(·). (right) Several common activation functions are shown.

Figure 3

Feed-forward, dense neural network: the input signals x_i are propagated forward through neurons which are arranged in layers of N_j elements as in Figure 2 until they reach the output neuron. The weights w_ij control the propagation of the signal through the neurons and layers.

Figure 4

Example of a convolutional neural network. It is composed of a convolutional layer, followed by a pool layer, and ending in a fully connected layer. Each layer can be understood as a matrix of weights.

Figure 5

Examples of more advanced neurons considering different topologies depending on the presence of forward or backward skip connections.

Figure 6

Example of Unet architecture. This is a complex network composed of different layers and connections, but it follows the architecture encoder-decoder.

Figure 7

Example of the GAN architecture. The generator attempts to produce images as similar as it is possible to the ground truth. A second network chooses between the restored image and the ground truth.

Figure 8

SPA workflow. The images are acquired as movies (frame collection). They are aligned to correct the beam-induced motion and averaged to reduce the noise variance. Then, the CTF is estimated to correct the microscope aberrations and defocus. Particles are selected to be later classified and screened in a 3D classification used to refine the structure. Finally, the map is sharpened to enhance the visualization, helping to build the atomic model (if it is possible).

Figure 9

Time evolution of the number of publications about movie alignment based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 10

Time evolution of the number of publications of CTF estimation based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 11

Time evolution of the number of publications of picking based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications

Figure 12

Time evolution of the number of publications on 2D classification based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications

Figure 13

Time evolution of the number of publications on reconstruction based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 14

Time evolution of the number of publications on initial volume algorithms based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 15

Time evolution of the number of publications about 3D classification based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 16

Time evolution of the number of publications about map refinement based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 17

Time evolution of the number of validation-related methods based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 18

Time evolution of the number of publications of resolution-related methods based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 19

Time evolution of the number of publications on map restoration based on analytical approaches or deep learning methods. The symbol #publications denotes the number of publications.

Figure 20

Time evolution of the number of publications on all workflow steps.