A Segmentation-aware Synergy Network for Single Particle Recognition in Cryo-EM

Abstract

Cryo-electron microscopy (cryo-EM) single particle analysis (SPA) has been an indispensable technology to reconstruct three-dimensional (3D) structures of biomolecules at near-atomic resolution. Tens of thousands of particles are required to obtain high-resolution 3D reconstructions, nevertheless, it is rather challenging due to the extremely noisy microscopy images and the diversity of particles. Recently, while deep learning-based methods have been devoted into the improvement of particle feature extraction and location estimation, most of them are plagued with vulnerable feature representation, inexact supervised ground truth. Furthermore, these DL-methods usually adopt denoising and particle picking as two-stage operations in the existing pipeline, which is inadequate to achieve accurate estimation for location. In this paper, we propose a segmentation-aware synergy framework to automatically select particles in which two tightly-coupled networks are designed including a multiple output convolution subnet for denoise to jointly learn strong object representation and pixel representation simultaneously and a deep convolution subnet for particle location. Furthermore, joint learning of the two networks can effectively enhance the synergy relationship between denoising and downstream recognition, thus leading to accurate and reliable location estimations for SPA. When applied with various EMPAIR real-world datasets, our model improves the performance of particle detection and exaction, especially intersection over union metric, and this strength has important implications for the next 2D alignment, 2D classification averaging, and high-resolution 3D refinement steps in SPA.