#31 · Structural bio & protein design

Cryo-EM and cryo-ET density guides generative and regression models of atomic structure

The dominant thread is experimental cryo-EM or cryo-ET density used as the data basis for learned models of molecular structure. Cryo-SWAN applies a wavelet-inspired voxel variational autoencoder to cryo-EM density volumes, improving reconstruction and enabling latent organization. DeepMDTOMO is a supervised 3D convolutional encoder-decoder that predicts all-atom coordinates from cryo-ET subtomograms as a faster alternative. A separate preprint maps noisy cryo-EM particle images to atomic coordinates through latent representations. MW-RaMViD adapts random-mask video diffusion to generate missing cryo-ET tilt images, with progressive sliding-window completion. What is new is the range of formulations, from variational latent models and direct regression to diffusion, applied to the same experimental observables. The thread is dominant rather than universal, since MW-RaMViD completes measurement data rather than predicting coordinates.