Over the last decade, cryo-electron microscopy (cryo-EM) has emerged as a powerful imaging technology for visualizing the 3D structure of proteins and other biomolecules at near-atomic resolution. Unlike other methods in structural biology, cryo-EM is uniquely poised to image large and dynamic protein complexes. However, this promise is limited by the computational task of 3D reconstruction, where a dataset of noisy and unlabeled 2D projection images are combined to infer the 3D structure(s) of the molecule of interest.
In this seminar, I will introduce cryoDRGN, a machine learning system for heterogeneous cryo-EM reconstruction. Underpinning the cryoDRGN method is a deep generative model parameterized by a new neural representation of 3D volumes and a learning algorithm combining exact and variational inference to optimize this model 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 to broaden the scope of cryo-EM to new classes of dynamic protein complexes. Finally, I will discuss how recent advances in machine learning for protein structure prediction (e.g. AlphaFold and Rosettafold) can complement methods for cryo-EM structure determination and what key algorithmic challenges remain to realize the next era of 3D visual biology.