Adaptive 3D Molecule Generation: Leveraging Structural Variations to Expand the Scope of Diffusion-Based Models

The proposed Geometric Adaptive Diffusion Model (GADM) can adaptively generate 3D molecules with desired structural variations, such as scaffold and ring-structure, by leveraging an equivariant masked autoencoder to capture the structural-grained representations as domain supervisors for the diffusion-based generation process.