An Unsupervised Classification Algorithm for Heterogeneous Cryo-EM Projection Images Based on Autoencoders.

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Heterogeneous three-dimensional (3D) reconstruction in single-particle cryo-electron microscopy (cryo-EM) is crucial for understanding the conformational dynamics of flexible biological macromolecules.
• Existing methods often rely on supervised learning or require extensive a priori knowledge, limiting their practical application.

Purpose of the Study:

• To develop an unsupervised algorithm for classifying heterogeneous cryo-electron microscopy (cryo-EM) projection images.
• To enable accurate 3D reconstruction of biological macromolecules without requiring labeled data or prior structural information.

Main Methods:

• An unsupervised heterogeneous cryo-EM projection image classification algorithm based on autoencoders was developed.
• Two autoencoder architectures (multi-layer perceptron and residual networks) were implemented to extract latent variables from projection images.
• Uniform Manifold Approximation and Projection (UMAP) was used for dimensionality reduction, followed by spectral clustering.

Main Results:

• The proposed algorithm effectively extracts category features from heterogeneous projection images.
• High classification and reconstruction accuracy were achieved on two heterogeneous cryo-EM datasets.
• The method demonstrates effectiveness in heterogeneous 3D reconstruction for single-particle cryo-EM.

Conclusions:

• The unsupervised autoencoder-based algorithm provides a powerful tool for analyzing conformational heterogeneity in cryo-EM.
• This approach overcomes limitations of supervised methods and a priori knowledge requirements.
• The algorithm facilitates accurate structural elucidation of flexible biological macromolecules.