Fig. 1: Schematic illustration of DynaMight.
From: DynaMight: estimating molecular motions with improved reconstruction from cryo-EM images
Two separate encoders take experimental images from each half set as input, and output a latent vector describing their conformational state. The decoders take the latent vectors together with the coordinates of Gaussian models for the consensus structures for each half set and generate a 3D deformation field for those Gaussians. The deformed models are then projected and compared to the experimental image in the loss function. At the end of the procedure, an approximation to the inverse deformation is used for reconstruction of an improved consensus map for each half set.
