ml-struct-bio/cryodrgn

What it solves

CryoDRGN solves the problem of heterogeneous reconstruction in cryo-electron microscopy (cryo-EM) and cryo-electron tomography (cryo-ET). Traditional methods often struggle to represent the continuous structural variations of biological molecules, whereas this tool allows for the modeling of a continuous distribution of 3D structures.

How it works

The project uses a neural network based representation for the volume. Specifically, it employs a Variational Autoencoder (VAE) to learn a latent variable representation of the structural heterogeneity. Users preprocess image stacks, parse poses and CTF parameters, and train the VAE to reconstruct 3D volumes from 2D projections.

Who it’s for

It is designed for structural biologists and researchers using cryo-EM and cryo-ET to study the complex, flexible biological molecules.

Highlights

• Continuous Distribution Modeling: Moves beyond discrete classes to model structural flexibility.
• Integrated Pipeline: Includes tools for downsampling, pose/CTF parsing, and voxel-based backprojection for verification.
• Interactive Visualization: Provides an end-to-end interactive dashboard web application for visualizing results.
• GPU Acceleration: Supports multi-GPU parallelization to speed up training for high-resolution images.