RED-ML

Motivation

‍Serial Crystallography (SX) in structural biology involves collecting diffraction data from numerous micro- or nano-crystals to assemble a complete dataset. Unlike traditional crystallography, which uses a single large crystal, SX collects diffraction patterns from many small, randomly oriented crystals and merges them to determine the 3D structure of macromolecules.

SX is valuable for studying proteins that don’t form large, high-quality crystals and for capturing ultrafast dynamic processes in time-resolved experiments. It involves illuminating numerous protein crystals with high-intensity X-ray beams at Free Electron Lasers (FEL) and synchrotrons, generating massive datasets of thousands to millions of diffraction patterns.

To obtain the final structure, each diffraction pattern must be indexed (i.e., determine crystal orientation) and merged into a complete dataset. Recent advancements in X-ray detectors, like the JUNGFRAU 4-megapixel (4M) detector, which operates at a maximum frame rate of 2kHz, have revolutionized macromolecular crystallography (MX). However, these advancements present challenges for real-time indexing algorithms and user feedback due to the enormous data volumes. Real-time feedback can help classify frames with valuable information, alleviating data storage issues.

Proposed Approach / Solution

‍We introduce the TORO (TOrch-powered Robust Optimization) indexer, a robust and adaptable indexing algorithm developed using the PyTorch framework. The proposed method relies on efficient sampling of the space of rotations, combined with robust regression with outlier rejection. For the robust regression part, we propose an annealing strategy to solve a sequence of robust estimation problems, akin to Trimmed Least Squares, until the error is below a prescribed threshold. The problem to solve is of combinatorial in nature, and the proposed method iteratively updates (in a greedy fashion) the support set of inlier measurements that are finally used for determining the crystal orientation. Besides state-of-the art estimation performance, the TORO indexer also enjoys nice computational and algorithmic features and is capable of operating on GPUs, CPUs, and other hardware accelerators supported by PyTorch, ensuring compatibility with a wide variety of computational setups. In our tests, TORO outpaces existing solutions indexing thousands of frames per second when running on GPUs, positioning it as an attractive candidate to produce real-time indexing and user feedback. Our algorithm streamlines some of the ideas introduced by previous indexers like DIALS real grid search and XGandalf, and refines them using faster and principled robust optimization techniques, which result in a concise codebase consisting of less than 500 lines.

Impact

‍Based on our evaluations across four proteins, TORO consistently matches and, in certain instances, outperforms established algorithms such as XGandalf and MOSFLM, occasionally amplifying the quality of the consolidated data while achieving indexing rates that are orders of magnitude higher. The inherent modularity of TORO and the versatility of Pytorch code bases, facilitate its deployment into a wide array of architectures, software platforms, and bespoke applications, highlighting its prospective significance in serial crystallography. The method has been successfully tested at both the Swiss Light Source (synchrotron) and Swiss FEL at PSI, as well as in other facilities in Europe including MAX IV, DESY, and ALBA.