igashov/DiffLinker

Equivariant diffusion architecture with SE(3)-invariant graph neural networks that learns to reverse a corruption process, generating linker atoms between molecular fragments without requiring attachment site specification. Supports both unconditional generation and protein-pocket conditioning via full atomic or backbone-only representations, with separate GNN auxiliary models for linker size prediction. Built on PyTorch Lightning, integrates with RDKit and OpenBabel for molecular I/O and validation across ZINC, GEOM, and PDB-derived datasets.

371 stars. No commits in the last 6 months.