DiffDesign–An SE(3)-Equivariant Diffusion Model for Molecule Linker Generation

Recent advances in targeted drug design have demonstrated that incorporating 3D structural information yields superior performance compared to target-free models, as this approach enables explicit modeling of atomic interactions in three-dimensional space. Generative models have achieved remarkable success in drug discovery, particularly in fragment-based drug design applications. In this work, we present DiffDesign, an SE(3)-equivariant three-dimensional conditional diffusion model designed to address the molecular linker design problem. Our model generates the missing atoms required to physically connect two disconnected molecular fragments, conditioned on a specific protein target. We demonstrate that DiffDesign outperforms existing methods on standard benchmark datasets, producing molecules with enhanced diversity and improved synthetic accessibility. Empirical studies validate that our model successfully generates chemically valid linkers with realistic three-dimensional structures when conditioned on target protein binding pockets.