Unsupervised Cryo-EM Images Denoising and Clustering Based on Deep Convolutional Autoencoder and K-Means+

Area of Science:

• Structural biology
• Biophysics
• Computational biology

Background:

• Cryo-electron microscopy (cryo-EM) is crucial for determining protein structures.
• Low signal-to-noise ratio (SNR) in cryo-EM images hinders accurate clustering and analysis.
• Existing methods struggle with the inherent noise in single-particle cryo-EM data.

Purpose of the Study:

• To develop an advanced method for denoising and clustering single-particle cryo-EM images.
• To improve the accuracy of structural determination by enhancing image quality.
• To address the challenge of low SNR in cryo-EM datasets.

Main Methods:

• An iterative denoising and clustering approach was developed.
• Key components include a deep convolutional variational autoencoder (DRVAE) for denoising and Balance size K-means++ (BSK-means++) for clustering.
• The method iteratively refines denoising using pseudo-supervision from reliable clustered samples.

Main Results:

• The proposed DRVAE with BSK-means++ method demonstrated superior denoising performance on single-particle cryo-EM images.
• Achieved higher clustering accuracy and normalized mutual information compared to existing methods.
• Generated reliable class-average images, preserving detailed information and avoiding class size imbalance.

Conclusions:

• The DRVAE with BSK-means++ method effectively improves the quality and clustering of cryo-EM images.
• This technique aids researchers in analyzing particle symmetry and heterogeneity.
• The approach enhances the reliability of structural determination in cryo-EM studies.