Cryo-SWAN: the Multi-Scale Wavelet-decomposition-inspired Autoencoder Network for molecular density representation of molecular volumes

A research team including Rui Li, Artsemi Yushkevich, Mikhail Kudryashev, and Artur Yakimovich has introduced Cryo-SWAN, a novel AI framework designed to address a critical gap in biomedical imaging. While most 3D computer vision methods work with point clouds or meshes, structural biology relies on volumetric density maps from techniques like cryo-electron microscopy (cryo-EM), which have been underexplored by AI. Cryo-SWAN is a voxel-based variational autoencoder that draws inspiration from multi-scale wavelet decomposition to learn robust representations of 3D molecular shapes directly from voxelized data, providing a practical, data-driven tool for the field.

The model's architecture performs conditional coarse-to-fine latent encoding and recursive residual quantization across different perception scales. This allows it to accurately capture both the global geometry and the high-frequency structural details present in complex molecular volumes. Evaluated on standard benchmarks like ModelNet40 and BuildingNet, as well as a newly curated cryo-EM dataset called ProteinNet3D, Cryo-SWAN consistently demonstrated superior reconstruction quality compared to existing 3D autoencoders. The learned latent space organizes molecules by shared geometric features, and the framework's integration with diffusion models enables powerful applications like volume denoising and conditional 3D shape generation, paving the way for new discoveries in structural biology.