Interpreting learned latent spaces of cryo-EM generative models

Abstract

Understanding the conformational heterogeneity of macromolecular complexes is a central challenge in structural biology. Recent advances in cryo-electron microscopy (cryo-EM) and deep generative models have enabled reconstruction of continuous distributions of 3D structures from single-particle images. However, interpreting the latent spaces of these models remains a major hurdle.

In this thesis, we explore methods for extracting biologically interpretable insights from trained cryo-EM generative models in a post hoc, model-agnostic manner. We first present an unsupervised framework which learns a volume-based embedding that enables continuous and discrete exploration of structural heterogeneity. We then introduce a minimally-supervised method that leverages prior biological knowledge of conformational dynamics to explicitly model specific, nonlinear motions, such as rigid-body rotations of protein subcomplexes.

Together, these methods demonstrate that cryo-EM generative models learn rich structural information that can be probed for meaningful biological signals.