Absorption correction for long-wavelength macromolecular crystallography

Thesis

This thesis presents novel methods for improving long-wavelength macromolecular crystallography (MX), focusing on analytical absorption correction through segmented tomography reconstruction. While the absorption effect is only a minor factor in standard macromolecular crystallography, it can become the largest source of uncertainty for experiments performed at long wavelengths. Current software packages for macromolecular crystallography typically employ empirical models to correct for the effects of absorption, with corrections determined by minimizing the differences in intensities between symmetry-equivalent reflections. These models are well-suited to capture smoothly varying experimental effects.

However, for very long wavelengths, empirical methods become an unreliable approach for modelling strong absorption effects with high fidelity. This issue is particularly acute when data multiplicity is low. This thesis addresses key challenges in absorption correction by introducing AnACor1.0, a ray-tracing analytical absorption correction method that utilizes segmented 3D models of crystal samples, including mounting loops and mother liquor. The accuracy of absorption correction is significantly improved compared to traditional empirical models, reducing systematic errors and enhancing data quality.

To further improve computational efficiency, AnACor2.0 was developed as a GPU-accelerated version of the absorption correction algorithm, leveraging CUDA-based implementation to achieve significant reductions in processing time. Additionally, CPU-based acceleration methods are introduced for cases with limited NVIDIA GPU resources. This acceleration makes the technique more practical for application to large-scale datasets, significantly reducing computational time and broadening its potential use in crystallography.

The combined contributions of this thesis significantly advance the field of long-wavelength macromolecular crystallography by improving the accuracy and efficiency of absorption correction methods. The development of GPU-accelerated algorithms, along with the integration of machine learning-based segmentation techniques, establishes a strong foundation for automated and rapid long-wavelength crystallographic data analysis. These advancements enable more precise experimental structural determinations in crystallography. Furthermore, when combined with AlphaFold (a protein structure predictive model that earned the Nobel Chemistry Prize 2024), these methods applied to real crystallography experiments provide opportunities for deeper insights into molecular structures, driving progress in structural biology and related fields.

Authors: Lu, Y

• DOI: 10.5287/ora-2rqvxwmay
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: scientific computing; X-ray crystallography; AI for science