Research On High Performance Cryo-electron Microscopy Three-Dimensional Reconstruction Techniques

With the rapid development of cryo-electron microscopy,the field of life sciences,especially its branch of structural biology,has received increasing attention.From the Nobel Prize in Chemistry awarded to cryo-electron microscopy in 2017 to the molecular architecture of the SARS-Co V-2 virus responsible for the COVID-19 pandemic resolved in 2020,human society has entered the century of life sciences.Structural biology in the field of life sciences has played a crucial part in drug design,vaccine development,and disease prevention and control.Although an increasing number of near-atomic resolution protein or biomolecular complex structures have been resolved,the large-scale application,the high-resolution analysis,and the automated application of cryo-electron microscopy technology still face many challenges.These challenges are mainly reflected in three aspects: firstly,the rapid increase in the number of images collected by cryo-electron microscopy and the increasingly complicated three-dimensional reconstruction algorithms have led to a significant increase in computational amount,making the required computation time unacceptable;secondly,the demand for the structure of proteins or biomolecular complexes in situ in cells has increased,but the resolution that can be achieved by cryo-electron tomography technology for reconstructing in situ structures cannot meet this demand;finally,the rapid increase in the number of images collected by cryo-electron microscopy has also increased the demand for automatic recognition of biomolecular particles from images.In response to these challenges,this dissertation aims to optimize cryo-electron microscopy threedimensional reconstruction image processing technology for computational efficiency,resolution,and automation,based on the mainstream single-particle three-dimensional reconstruction software THUNDER in the field of structural biology and the GPU platform.The main research content and innovations of this article are as follows:1.Designing a GPU parallel computing framework for THUNDER.This dissertation used optimization methods such as overlapping computation and communication,algorithm logic optimization and algorithm fine-grained parallelism,GPU memory optimization and mixed-precision optimization to achieve efficient utilization of GPU memory resources and effective scheduling of computing resources.These optimization strategies have enabled the THUNDER-based GPU parallel computing framework designed in this dissertation to achieve a fine speed-up ratio and scalability.2.Designing a tilt image signal de-overlapping algorithm for the subtomogram averaging step in cryo-electron tomography,which can effectively remove non-target signals that obstruct the target particle signal in the tilted image.Moreover,this dissertation extended the single-particle 3D reconstruction software THUNDER by adding subtomogram averaging and signal de-overlapping algorithms,to achieve high-resolution subtomogram averaging.3.Designing an automated biological macromolecule particle picking algorithm framework based on meta-learning.This dissertation used the meta-training step of joint data sets composed of particle images with different structural features and the parameter fine-tuning step based on real data sets to enable the parameter model in the algorithm framework to have good generalization performance and high accuracy in macromolecule particle recognition.