Abstract
Structure-based drug design (SBDD), which aims to generate molecules that canbind tightly to the target protein, is an essential problem in drug discovery,and previous approaches have achieved initial success. However, most existingmethods still suffer from invalid local structure or unrealistic conformationissues, which are mainly due to the poor leaning of bond angles or torsionalangles. To alleviate these problems, we propose AUTODIFF, a diffusion-basedfragment-wise autoregressive generation model. Specifically, we design a novelmolecule assembly strategy named conformal motif that preserves theconformation of local structures of molecules first, then we encode theinteraction of the protein-ligand complex with an SE(3)-equivariantconvolutional network and generate molecules motif-by-motif with diffusionmodeling. In addition, we also improve the evaluation framework of SBDD byconstraining the molecular weights of the generated molecules in the samerange, together with some new metrics, which make the evaluation more fair andpractical. Extensive experiments on CrossDocked2020 demonstrate that ourapproach outperforms the existing models in generating realistic molecules withvalid structures and conformations while maintaining high binding affinity.