Structural Similarity Based Low SNR Single Particle Cryo-EM Images Reconstruction

Cryo-EM 3D reconstruction technology has a rapid development in biomolecule's 3D structure study in recent years.Compared to classic structural biology methods,it requires no crystallization and fewer restrictions on samples and is highly efficient with fast image processing tools.It attracts more and more researchers' attention.Usually a low electron dose will be used to avoid radiation damage on the targeting molecules.And that makes microscopy images exhibit very low Signal-to-Noise-Ratio(SNR),which is the main obstacle for cryo-EM image reconstruction.Single particle image reconstruction mainly has three steps,image clustering,initial model generation and model refinement.All these steps use clustering method to refrain noise and get clear class averages.The key component of clustering method is the measurement of similarity between two images.Currently,popular single particle reconstruction methods usually directly measure similarity between two images,only consider information of two images and nothing more,which will be easily biased by the hidden noise.In this paper,we propose a network based single particle image reconstruction method.We construct a network to calculate image similarities.Each node represents an image in dataset.And then we use network's topology to refine similarity between two images and construct a refined network,which convert the clustering problem into network partition problem.The refined similarity is more robust to noise and is more precise in assigning particle's class.Our experimental results on both synthetic and real datasets demonstrate its advantages over current reconstruction methods.We also write a software implementing single particle reconstruction procedure.We adopt a new resolution calculation method,making it possible to automatically select initial models and reconstruction.Our software needs very few priori knowledge and manual intervention.