Alfredo Bellisario: Deep learning assisted phase retrieval and computational methods in coherent diffractive imaging

Abstract

In recent years, advances in Artificial Intelligence and experimental techniques have revolutionized the field of structural biology. X-ray crystallography and Cryo-EM have provided unprecedented insights into the structures of biomolecules, while the unexpected success of AlphaFold has opened up new avenues of investigation. However, studying the dynamics of proteins at high resolution remains a significant obstacle, especially for fast dynamics. Single-particle imaging (SPI) or Flash X-ray Imaging (FXI) is an emerging technique that may enable the mapping of the conformational landscape of biological molecules at high resolution and fast time scale. This thesis discusses the potential of SPI/FXI, its challenges, recent experimental successes, and the advancements driving its development. In particular, machine learning and neural networks could play a vital role in fostering data analysis and improving SPI/FXI data processing. In Paper I, we discuss the problem of noise and detector masks in collecting FXI data. I simulated a dataset of diffraction patterns and used it to train a Convolutional Neural Network (U-Net) to restore data by denoising and filling in detector masks. As a natural continuation of this work, I trained another machine learning model in Paper II to estimate 2D protein densities from diffraction intensities. In the final chapter, corresponding to Paper III, we discuss another experimental method, time-resolved Small Angle X-ray Scattering (SAXS), and a new algorithm recently developed for SAXS data, the DENsity from Solution Scattering (DENSS) algorithm. I discuss the potential of DENSS in time-resolved SAXS and its application for structural fitting of AsLOV2, a Light-Oxygen-Voltage (LOV) protein domain from Avena sativa.