Diff-Shape: A Novel Constrained Diffusion Model for Shape based De Novo Drug Design

Shape-based virtual screening is a widely utilized method in ligand-based de novo drug design, aiming to identify molecules in chemical libraries that share similar 3D shapes but simultaneously possess novel 2D chemical structures compared to the reference compound. As an emerging technology, generative model is an alternative way to do de novo drug design by directly generating 3D novel structures. However, existing models face challenges in reliably generating valid drug-like molecules under specific conformation constrains. Here, a novel diffusion model constrained with 3D reference shape, Diff-Shape, was proposed to generate structures whose 3D conformations are similar to a given reference shape, thereby avoiding the computational cost of screening large database of 3D conformations. This model utilized a zero-weighted graph control module, taking in various forms of point clouds of reference shape to guide diffusion process of 3D molecular generation. The results show that our model is capable of generating molecules with high shape similarity but still low 2D graph similarity to the query structure and it significantly out-performs existing shape based generative models. Examples were also given to demonstrate that combining with docking software our model can effectively generate structures in structure-based drug design scenario.