Research On Particle Picking In Cryo-electron Micrographs Based On URDnet

Particle picking is a crucial step for single-particle cryo-electron microscopy and the first step for 3D reconstruction of single-particle biomacromolecules.The quality of the selected particles directly affects the efficiency of the single-particle macromolecular reconstruction and final resolution of 3D structure determination.The low signal-to-noise ratio,low contrast,and severe background noise,and many impurities in cryo-EM images make the particle recognition difficult to ensure efficiency and reliability.High-resolution 3D reconstruction of cryo-electron microscopy images usually requires tens of thousands of particles.Picking particles from images is time-consuming and laborious and has become the practical bottleneck for the determination of biomacromolecules structures in Cryo-EM.For this problem,we propose a U-Net-based residual dense neural network(URDnet)for accurate and automatic particle localization from cryo-EM images.The content of the proposed method is summarized as follows:1.In view of the low signal-to-noise ratio and uneven background intensity of cryo-EM images,the particle recognition accuracy in the image is quite low.The paper introduces data preprocessing to improve the quality of Cryo-EM images.Apply histogram equalization and Wiener filtering algorithms to increase contrast and reduce noise.As a result,the pre-processed image is more conducive to the feature extraction of the Cryo-EM image by the designed network.The designed network selects the popular deep learning network U-Net in image semantic segmentation as the basic network framework to achieve global feature fusion so that particles can be accurately detected from background noise with a small number of training images.Embedding the residual dense block in the encoder part of the U-Net architecture not only realizes the local residual feature learning but also enhances the local dense feature fusion,thereby further improving the precision of artifacts segmentation in Cryo-EM images.2.Aiming at the burden of manual labeling caused by the large number of particles in the Cryo-EM image,and the interference of carbon film,ice slag,dissociated particles,stacked particles and other impurities on particle picking,an annotation method combining point-level labels and pixel-level labels was proposed,using different levels of supervision for different objects in the Cryo-EM images,which greatly improves the efficiency of image labeling and avoids selecting particles from the impurity region.Moreover,connected component analysis was introduced in the particle picking step to obtain the information of all connected domains labeled as particles to locate the candidate particles.Some thresholds are set by the parameters such as the mean area of all single connected components to eliminate dissociated or stacked particles,which improves the accuracy of picked particle greatly.The method trained and tested the public dataset of 80 S ribosomes,HCN1 channel,Tcd A1 toxin subunit and KLH.The experimental results showed our method reached excellent particle picking performance and high applicability to multiple protein data,and acquire the 3D structure of picked particles with higher resolution compared with other mainstream methods.