Deep learning of elastic 3D shapes for cryo electron microscopy analysis of continuous conformational changes of biomolecules

Cryo electron microscopy (cryo-EM) allows high-resolution 3D reconstruction of biomolecular structures from highly noisy 2D parallel-beam projection images containing tens of thousands of copies of the same macromolecular complex but at different random orientations and positions. However, biomolecular complexes are not rigid but flexible entities that change their conformations gradually (continuous transition with many intermediate states) to accomplish biological functions (e.g., DNA replication, protein synthesis, etc.). The determination of the full distribution of conformations (conformational space or landscape) from cryo-EM images is challenging but could provide insights into working mechanisms of the complexes. In this paper, we present a method for conformational space determination, which uses deep learning in combination with cryo-EM image analysis and normal mode analysis (molecular mechanics simulation), where the amplitudes of normal modes are used as parameters of the elastic 3D shapes of complexes (the parameters determining the conformation). We show the performance of this new method using synthetic cryo-EM data.