Towards Fast and Robust Algorithms in Flash X-ray single-particle Imaging

Modern X-ray Free Electron Laser (XFEL) technology provides the possibility to acquire a large number of diffraction patterns from individual biological nano-particles, including proteins, viruses, and DNA. Ideally, the collected data frames are high-quality single-particle diffraction patterns. However, unfortunately, the raw dataset is noisy and also contains patterns with scatterings from multiple particles, contaminated particles, etc. The data complexity and the massive volumes of raw data make pattern selection a time-consuming and challenge task. Further, X-rays interact with particles at random and the captured patterns are the 2D intensities of the scattered waves, i.e. we cannot observe the particle orientations and the phase information from the 2D diffraction patterns. To reconstruct 2D diffraction patterns into 3D structures of the desired particle, we need a sufficiently large single-particle-pattern dataset. The computational methodology for this reconstruction task is still under development and in need of an improved understanding of the algorithmic uncertainties. In this thesis, we tackle some of the challenges to obtain 3D structures of sample molecules from single-particle diffraction patterns. First, we have developed two classification methods to select single-particle diffraction patterns that are similar to provided templates. Second, we have accelerated the 3D reconstruction procedures by distributing the computations among Graphics Processing Units (GPUs) and by proposing an adaptive discretization of 3D space. Third, to better understand the uncertainties of the 3D reconstruction procedure, we have evaluated the impact of the different sources of resolution-limiting factors and introduced a practically applicable computational methodology in the form of bootstrap procedures for assessing the reconstruction uncertainty. These technologies form a data-analysis pipeline for recovering 3D structures from the raw X-ray single-particle data, which also analyzes the uncertainties. With the experimental developments of the X-ray single-particle technology, we expect that the data volumes will be increasing sharply, and hence, we believe such a computational pipeline will be critical to retrieve particle structures in the achievable resolution