Parallel Implementation Of Single Particle Cryo-EM Structure Determination On GPU

Single-particle cryo-electron microscopy three-dimensional reconstruction is one of the most promising techniques in structural biology.In order to obtain a high-resolution reconstruction model,we need to deal with the images of the cryo-electron microscopic images,which need powerful computing power.Traditional serial programs and multicore CPU programs can not meet the needs of scientists in the field.RELION is currently the most popular cryomicroscope image processing software,but its computing needs are currently frozen electron microscope software,the largest.In this paper,based on the 3D reconstruction algorithm of single particle cryo-electron microscopy and the RELION program,the 3D reconstruction of single particle cryo-electron microscopy is designed and implemented by GPU and CUDA parallel optimization technique.The main work of this paper are as follows:1)Analyze and study the mathematics foundation,correlation principle and Bayesian EM algorithm of three-dimensional reconstruction of single particle cryoelectron microscopy,and analyze the hot and characteristic of RELION program.2)According to the program features of RELION,a GPU-based multi-stage parallel single-particle cryo-electron microscopy image processing framework is designed,which adapts the GPU structure from cooperative computation,task partitioning and storage management.Through cyclic segmentation,Structure is reconstructed,and an adaptive parallel framework is proposed to avoid the problem that the storage space is not enough.3)Based on GeRelion parallel framework,the program model of multi-loop,nondeterministic threshold calculation,global protocol operation and intensive memory access operation is optimized and optimized.A rough parallel granularity and high parallelism are combined to optimize the cycle.Non-deterministic threshold computation optimization,global protocols based on on-chip atomic operations,and data-sharing methods based on on-chip memory.The experimental results show that the GPU-based multi-stage parallel freezeelectron microscope image processing software can achieve more than 60 times faster acceleration,and the multi-level parallel freeze-electron microscope image processing software can achieve more than 60 times acceleration,and in multi-data set configuration methods GeRelion,GPU has a very good scalability.