MolBFN: Equivariant Bayesian Flow Networks for 3D Molecule Generation

This work introduces MolBFN, a new generative SE(3)-equivariant model utilizing Bayesian Flow Networks (BFNs), which models the parameters of molecule distribution in a continuous and differentiable way. Unlike most of previous models adding chemical bonds via predefined rules, MolBFN integrates dynamic interactions between atom and chemical bonds by utilizing an innovative graph neural network which updates both atom and bond representations during message passing simultaneously. MolBFN unifies the probabilistic distribution modeling across various modalities of atoms and bonds by maintaining a coherent and continuous representation between atoms and bonds while preserving SE(3)-equivariance. Experimental results demonstrate that MolBFN can generate molecules with significantly better validity and stability scores.