Study On Single Particle Cryo-EM Image Classification Based On Neural Network And Density Peaks Clustering

Cryo-electron microscopy(Cryo-EM)has achieved fruitful results in the field of structural biology in recent years.Among them,single-particle Cryo-EM image classification plays an important role in improving the resolution of three-dimensional model reconstruction.There is still much room for improvement of the existing PCA-based methods because of the low signal-noise ratio(SNR)of single-particle images.However,there is no research on the application of new techniques such as neural networks to solve the single-particle image classification problem.In this paper,we take single particle image processing as the research background,to explore the application of neural network and density peak clustering algorithm in the single particle Cryo-EM image classification.This paper focuses on the application of neural network feature extraction of particle images,and optimization of density peak clustering algorithm to classify the image features.Neural network method to extract particle image features face the following problems: low signal-to-noise ratio of particle image and insignificant features result in the ineffectively application of existing deep learning methods;image has rotation transform,but neural network usually can not extract the main features of particles correctly;due to the diversity of particle data,particle images are often unlabeled,supervised neural network classification method is not suitable for particle image classification problem.In particle image classification,due to the high dimensionality of the particle features,the density peaks clustering algorithm is better than K-means,but there is a problem of high computational complexity in the particle image data.This article studies the problems that these two methods face,the main work is as follows:Two kinds of single particle Cryo-EM feature extraction methods based on neural network are proposed.One is the combination of SIFT feature and autoencoder(SIFT +AE).Experiments show that the classification accuracy of 93% can be achieved,which is better than the results of using SIFT alone or pure neural network such as DEC(Deep Embedding Clustering),DSC(Deep Subspace Clustering)to extract features.In addition,this paper proposes a single-layer convolutional autoencoder to extract the features and classify the particle images by the bag-of-words model.Although the effect of the SIFT+ AE method is not achieved,it achieves the best accuracy and NMI in several simple neural network feature extraction methods.High Efficiency Density Peak Clustering Parallel Algorithm Based on GPU.This paper analyse the principle of the algorithm to locate its computational bottlenecks,and evaluate its potential of parallelism.In the light of the analysis,we propose CUDADP,an efficient parallel algorithm of DP targeting on GPU architecture,and implement this parallel method with CUDA.Specifically,we use shared memory to improve the data locality,which reduces the amount of global memory access.In order to exploit the coalescing accessing mechanism of GPU,we reconstruct the data structure of the CUDADP program from AOS(Array of Structure)to SOA(Structure of Array).In addition,we introduce the binary search and sampling method to avoid sorting large array.The experimental results show that CUDA-DP classification accuracy can reach 95%,better than K-means classification results,and CUDA-DP achieves a speedup of over 45 times when compared to the CPU-based density peaks implementation in single particle CryoEM image classification tasks.In summary,this paper studies and proposes a single particle cryo-EM image classification method based on neural network and GPU-based high-density peak clustering parallel algorithm,Experimental results show that the classification method proposed in this paper is more accurate than the other neural network-based methods in single particle Cryo-EM image classification,and the CUDA-DP proposed in this paper can obtain a very good acceleration effect and is successfully applied in the single particle Cryo-EM image classification issues.