Deep learning methods to study structurally heterogeneous macromolecules in vitro and in situ

Abstract

Proteins, RNA, and other biomolecules form complex 3-D structures that dynamically interact to carry out essential biological processes. These macromolecular complexes are often structurally heterogeneous, which is key to executing or regulating their specific biological functions.

To understand the molecular mechanisms underpinning these biological functions, structural biologists aim to determine the 3-D structure of the relevant macromolecule or macromolecular complex. Most such structural insights use techniques that strip the macromolecule of its cellular context (i.e., in vitro) and, subsequently, report a single average structure. However, recent advances in cryogenic electron microscopy (cryo-EM) provide avenues to determine sets of heterogeneous structures from a single dataset, and simultaneous advances in cryogenic electron tomography (cryo-ET) enable the resolution of macromolecules in their native cellular environment (i.e., in situ).

This thesis describes the conceptualization, implementation, and application of tomoDRGN, a deep learning method developed to resolve structurally heterogeneous macromolecules in situ. TomoDRGN builds on the well characterized cryoDRGN method, which facilitates analysis of heterogeneous structures by cryo-EM, to cryo-ET, where I show it efficiently learns an ensemble of unique 3-D volumes from the structurally heterogeneous dataset provided. I additionally describe the application of TomoDRGN to datasets of diverse macromolecules, highlighting its ability to resolve conformational and compositional heterogeneity and to identify rare yet biologically informative structural states. This thesis also details an approach and protocol for rapid structural characterization of bacterial ribosomes in situ, wherein tomoDRGN facilitates powerful upstream dataset filtration. Finally, this thesis provides a detailed protocol for the characterization of heterogeneous cryo-EM datasets with cryoDRGN and, in doing so, illustrates the types of new insights enabled by the cryoDRGN and tomoDRGN Deep Reconstructing Generative Networks.