This work is based on the manifold-embedding approach to study biological molecules exhibiting continuous conformational changes. Previous work established a method capable of reconstructing 3D movies and accompanying energetics of atomic-level structures from single-particle cryo-EM images of macromolecules displaying multiple conformational degrees of freedom. Here, we introduce an unsupervised geometric machine learning approach that is informed by detailed heuristic analysis of manifolds formed by simulated heterogeneous cryo-EM datasets generated from an atomic structure. These simulated data were generated with increasing complexity to account for multiple conformational motions, state occupancies and typical microscope parameters in a wide range of signal-to-noise ratios. Using these datasets as ground-truth, we provide detailed exposition of our findings using several conformational motions while exploring the available parameter space. Guided by these insights, we build a framework to leverage the high-dimensional geometric information obtained towards reconstituting a quasi-continuum of conformational states in the form of a free-energy landscape and respective 3D density maps for all states therein. As shown by a direct comparison of results, this framework offers substantial improvements relative to the previous work.

Geometric machine learning informed by ground truth: Recovery of conformational continuum from single-particle cryo-EM data of biomolecules