TORO Indexer: a PyTorch-based indexing algorithm for kilohertz serial crystallography

Area of Science:

• Structural Biology
• Computational Crystallography
• Biophysics

Background:

• Serial crystallography (SX) is crucial for determining protein structures using X-ray diffraction.
• Processing SX data involves indexing diffraction patterns to determine crystal orientation, integration, and merging.
• Existing indexing methods can be computationally intensive, limiting real-time analysis.

Purpose of the Study:

• To introduce TORO (Torch-powered robust optimization) Indexer, a novel algorithm for accelerating SX data processing.
• To develop a robust and adaptable indexing solution leveraging the PyTorch framework.
• To improve the efficiency and accuracy of crystal orientation determination in SX.

Main Methods:

• Development of the TORO Indexer using the PyTorch deep learning framework.
• Implementation of robust optimization techniques for efficient indexing.
• Benchmarking TORO's performance against established algorithms like XGandalf and MOSFLM on protein datasets.

Main Results:

• TORO demonstrates significantly faster indexing speeds, processing thousands of frames per second on GPUs.
• The algorithm achieves comparable or superior data quality to existing methods.
• TORO's performance was evaluated across four different protein datasets, showing consistent results.

Conclusions:

• TORO Indexer offers a substantial speed improvement for SX data processing.
• Its adaptability and performance make it a valuable tool for real-time analysis and structure determination.
• The PyTorch-based design ensures broad compatibility and potential for integration into various computational workflows.