arneschneuing/DiffSBDD

Implements equivariant graph neural networks to maintain SE(3) symmetry during diffusion sampling, enabling generation of both C-alpha backbones and full-atom ligand structures conditioned on protein binding pockets. Supports multiple design tasks including de novo molecule generation, substructure inpainting for scaffold elaboration, and evolutionary optimization via noising-denoising cycles. Built on PyTorch Lightning with pre-trained checkpoints on CrossDocked and Binding MOAD benchmarks, integrating RDKit and BioPython for molecular and structural processing.

488 stars. No commits in the last 6 months.