Research And Application Of Cryo-EM 3D Reconstruction Algorithm Based On Convolution Theory

Cryo-EM is a technique for 3D reconstruction of biological samples,which enables the acquisition of 3D structures at near-atomic resolution.With the increase of data volume and the improvement of image resolution,people have higher demands on the speed and efficiency of algorithms.In order to solve the growing demand and facilitate more research,reconstruction algorithm software packages such as THUNDER have been developed.THUNDER has the characteristics of adaptive parameter adjustment,which can reduce the difficulty of 3D reconstruction.However,THUNDER’s convolution algorithm relies on FFT and i FFT,resulting in poor program parallelism.So it makes sense to further optimize THUNDER.This paper mainly optimizes THUNDER’s convolution algorithm.Firstly,the 3D circular convolution is improved from two aspects of memory space and number of cycles;secondly,the improved algorithm is embedded in THUNDER,which not only maintains the resolution of the 3D reconstruction structure,but also successfully realizes the parallel calculation of convolution;Finally,the improved algorithm is implemented in the GPU and adapted to the situation of multiple GPUs.The calculation speed is about 3.42 times higher than that of a single GPU.This paper specifically conducts the following four parts of the study.1.Study the optimization scheme of the convolution calculation part of the THUNDER program.The research plan is to use the convolution theorem to replace the method of calculating convolution using FFT in the THUNDER program with the method of directly calculating convolution.The 1D and 3D discrete frequency domain convolution theorems are introduced,and the feasibility of the optimization scheme is verified theoretically.In addition,the theory of the optimization scheme was verified,and the specific optimization scheme was implemented.2.An improved algorithm of 3D circular convolution is proposed.The steps of improving the algorithm are given and programmed by MATLAB,and the obtained results are consistent with those realized by the convolution theorem.Compared with the traditional algorithm,the improved algorithm only needs smaller memory space and fewer loops to get the same result.3.Embed the improved algorithm into the THUNDER program.In the Sunway supercomputing platform,the parallel computing program is implemented for the improved algorithm,and the necessary optimization is performed on the program.At the same time,it was embedded in the THUNDER program,and the modified program was run to obtain a 3D reconstructed structure.4.The improved algorithm is implemented in the GPU,the program is extended to the case of multiple GPUs operation,and the speed increasing about 3.42 times.