Accounting Conformational Dynamics into Structural Modeling Reflected by Cryo-EM with Deep Learning

Abstract

With the continuous development of structural biology, the requirement for accurate threedimensional structures during functional modulation of biological macromolecules is increasing. Therefore, determining the dynamic structures of bio-macromolecular at high resolution has been a highpriority task. With the development of cryo-electron microscopy (cryo-EM) techniques, the flexible structures of biomacromolecules at the atomic resolution level grow rapidly. Nevertheless, it is difficult for cryo-EM to produce high-resolution dynamic structures without a great deal of manpower and time. Fortunately, deep learning, belonging to the domain of artificial intelligence, speeds up and simplifies this workflow for handling the high-throughput cryo-EM data. Here, we generalized and summarized some software packages and referred algorithms of deep learning with remarkable effects on cryo-EM data processing, including Warp, user-free preprocessing routines, TranSPHIRE, PARSED, Topaz, crYOLO, and self-supervised workflow, and pointed out the strategies to improve the resolution and efficiency of three-dimensional reconstruction. We hope it will shed some light on the bio-macromolecular dynamic structure modeling with the deep learning algorithms.

Keywords: Cryo-EM; Dynamic structures of biomacromolecules; data preprocessing; deep learning; particle selection; topazdenoise.