Generative Models

CryoLens is a 3D variational autoencoder for learning interpretable representations of cellular structures from cryo-electron tomography data (available on the CZI Virtual Cell Platform). The model learns compressed representations of 3D subtomograms that capture structural variation in an interpretable latent space—useful for visualization, classification, and understanding structural diversity. Trained on a large synthetic dataset composed of 100+ diverse structures, CryoLens can produce high-fidelity single-angle 3D densities. Tools are available for napari and the Virtual Cell Platform.

Nov 5, 2025