Analysis by Compression: An end-to-end approach to cryo-EM
Jasenko Zivanov

• 2021-09-13

Abstract
A novel approach to cryo-EM will be proposed which interprets the problem as one of compression. Instead of sequentially detecting, classifying and refining particles in cryo-EM micrographs, the analysis-by-compression approach instead aims to compress each micrograph into a controlled form using variational auto encoders. Since the noise in the micrographs is non-deterministic, any successful compression process is expected to encode the underlying physical reality into the controlled latent space. By choosing an appropriate form for the latent space, this encoding becomes interpretable. For regular single-particle data sets, the latent space corresponds to a set of points in a plane, each associated with a spatial orientation. This approach is expected to work on a wide range of cryo-EM modalities with minimal human supervision, and it holds the promise to allow solving smaller molecules than the traditional approach. Applications to different modalities will be discussed, including regular single-particle analysis, diffractive imaging, electron cryo-tomography and helical-filament processing.