Supervised classification methods for flash X-ray single particle diffraction imaging.

Area of Science:

• Biophysics
• Structural Biology
• Computational Imaging

Background:

• Modern X-ray Free Electron Lasers (XFELs) enable Flash X-ray single-particle diffraction Imaging (FXI) to capture millions of diffraction patterns daily.
• FXI data often contains noise from sample contamination and inherent biomolecular heterogeneity, complicating analysis.
• Efficiently reducing data volume and selecting high-quality patterns are crucial for FXI experiments.

Purpose of the Study:

• To develop and evaluate supervised machine learning methods for classifying FXI patterns.
• To address the challenges of data processing and quality control in high-throughput FXI experiments.
• To enable rapid, on-site classification of single-molecule diffraction patterns.

Main Methods:

• Implementation of two supervised template-based learning algorithms: Eigen-Image and Log-Likelihood classifiers.
• Development of methods for rapid template matching of single-molecule patterns.
• Design for straightforward parallelization to match XFEL repetition rates for real-time processing.

Main Results:

• The Eigen-Image and Log-Likelihood classifiers achieve millisecond-level pattern matching.
• Methods demonstrate stable performance across diverse synthetic datasets.
• Accurate classification of a real mimivirus dataset with 0.9 accuracy was obtained.

Conclusions:

• Supervised template-based learning offers an effective solution for classifying FXI patterns.
• The developed methods significantly enhance the efficiency and throughput of biomolecular imaging with FXI.
• On-site, real-time data processing is feasible, accelerating scientific discovery in structural biology.