Abstract

Machine learning for reconstructing dynamic protein structures from cryo-EM images

Major technological advances in cryo-electron microscopy (cryo-EM) have produced new opportunities to study the structural heterogeneity and dynamics of macromolecular complexes. However, this structural heterogeneity complicates 3D reconstruction and is traditionally addressed with discrete clustering approaches that fail to capture the full range of biomolecular dynamics. In this talk, I will overview cryoDRGN, a heterogeneous reconstruction algorithm that leverages the expressive representation power of deep neural networks to reconstruct continuous distributions of cryo-EM density maps. Underpinning the cryoDRGN method is a deep generative model parameterized by a new neural representation of 3D volumes and a learning algorithm to optimize this representation from unlabeled 2D cryo-EM images. Extended to real datasets and released as an open-source tool, cryoDRGN has been used to discover new protein structures and visualize continuous trajectories of their motion. I will discuss various extensions of the method for broadening the scope of cryo-EM to new classes of dynamic protein complexes, and analyzing the learned generative model. CryoDRGN is open-source software freely available at http://cryodrgn.csail.mit.edu.