kyegomez/Open-AF3

Replaces the EvoFormer with a 48-block PairFormer architecture that processes paired and single token representations, then passes them to a diffusion module that operates directly on atomic coordinates. Employs a denoising diffusion strategy where the network learns to predict clean atomic positions from progressively noised structures, with cross-distillation from AlphaFold2 predictions to mitigate hallucination, and confidence heads that predict per-atom/pairwise errors via LDDT and distance error metrics. Handles diverse biomolecular inputs including protein sequences, ligand SMILES, and covalent modifications while supporting structures up to thousands of residues.

802 stars.