Machine Learning for Reconstructing Molecular Flexibility in Cryo-
Electron Microscopy
Abstract:
Cryogenic electron microscopy (cryo-EM) is a powerful technique to
obtain the 3D structure of macromolecules from thousands of noisy
projection images. Since these macromolecules are flexible by nature,
the areas shows lower resolution and gives a blurry reconstruction. We
propose a novel method incorporated in a software package named
dynamight, that represents the molecule with gaussian basis functions
and estimates deformation fields for every experimental image by a
variational autoencoder. We further use the estimated deformations to
better resolve the flexible regions in the reconstruction using a
filtered backprojection algorithm along curved lines. We present results
on real data showing that we obtain improved 3D reconstruction.