Unsupervised learning approaches to characterizing heterogeneous samples using X-ray single-particle imaging

Area of Science:

• * X-ray single-particle imaging (SPI)
• * Structural biology
• * Materials science

Background:

• * Classifying structural heterogeneity in X-ray SPI is challenging due to low signal-to-noise ratios and orientation variations.
• * Existing methods struggle with diverse structural landscapes and complex datasets.

Purpose of the Study:

• * To develop robust methods for classifying structural heterogeneity in X-ray SPI data.
• * To enable the study of complex structural landscapes beyond homogeneous sample sets.
• * To advance the analysis of nanocrystal growth, dynamics, and phase transitions.

Main Methods:

• * Common-line principal component analysis (PCA) for parameter-free, automatic rough classification.
• * Variation auto-encoders (VAEs) for generating 3D structures across the structural landscape.
• * Integration with the noise-tolerant expand-maximize-compress (EMC) algorithm.

Main Results:

• * Demonstrated utility on experimental gold nanoparticle data with low photon counts per pattern.
• * Successfully recovered both discrete structural classes and continuous deformations.
• * Showcased robustness in handling orientation-induced variations and low signal-to-noise ratios.

Conclusions:

• * The developed methods offer a significant advancement over previous approaches for analyzing SPI data.
• * Opens new avenues for studying dynamic processes like nanocrystal growth and phase transitions.
• * Enables deeper insights into the structural landscape of various sample ensembles.