Abstract
Generating molecules that bind to specific proteins is an important but challenging task in drug discovery. Most previous works typically generate atoms autoregressively, with element types and 3D coordinates of atoms generated one by one. However, in real-world molecular systems, interactions among atoms are global, spanning the entire molecule, leading to pair-coupled energy function among atoms. With such energy-based consideration, modeling probability should rely on joint distributions rather than sequential conditional ones. Thus, the unnatural sequential auto-regressive approach to molecule generation is prone to violating physical rules, yielding molecules with unfavorable properties. In this study, we propose DiffBP, a generative diffusion model that generates molecular 3D structures, leveraging target proteins as contextual constraints at the full-atom level in a non-autoregressive way. Given a designated 3D protein binding site, our model learns to denoise both element types and 3D coordinates of an entire molecule using an equivariant network. In experimental evaluations, DiffBP demonstrates competitive performance against existing methods, generating molecules with high protein affinity, appropriate molecule sizes, and favorable drug-like profiles. Additionally, we developed a website server for medicinal chemists interested in exploring the art of molecular generation, which is accessible at https://www.manimer.com/moleculeformation/index.