microsoft/foldingdiff

Represents protein backbones as dihedral angles and uses a BERT-based diffusion model with cosine noise scheduling to iteratively denoise corrupted angle sequences into valid structures. Built on PyTorch Lightning with HuggingFace transformers, the framework supports both training on CATH dataset and sampling novel backbones via reverse diffusion, with utilities for PDB conversion, designability evaluation via inverse folding, and TM-score comparison against training data.

565 stars. No commits in the last 6 months.