DecompOpt: Controllable and Decomposed Diffusion Fashions for Construction-based MolecularOptimization
Authors: Xiangxin Zhou, Xiwei Cheng, Yuwei Yang, Yu Bao, Liang Wang, Quanquan Gu
Summary: Just lately, 3D generative fashions have proven promising performances in structure-based drug design by studying to generate ligands given goal binding websites. Nevertheless, solely modeling the target-ligand distribution can hardly fulfill one of many primary objectives in drug discovery — designing novel ligands with desired properties, e.g., excessive binding affinity, simply synthesizable, and so on. This problem turns into notably pronounced when the target-ligand pairs used for coaching don’t align with these desired properties. Furthermore, most current strategies intention at fixing textit{de novo} design activity, whereas many generative eventualities requiring versatile controllability, akin to R-group optimization and scaffold hopping, have obtained little consideration. On this work, we suggest DecompOpt, a structure-based molecular optimization technique based mostly on a controllable and decomposed diffusion mannequin. DecompOpt presents a brand new era paradigm which mixes optimization with conditional diffusion fashions to attain desired properties whereas adhering to the molecular grammar. Moreover, DecompOpt presents a unified framework overlaying each textit{de novo} design and controllable era. To realize so, ligands are decomposed into substructures which permits fine-grained management and native optimization. Experiments present that DecompOpt can effectively generate molecules with improved properties than robust de novo baselines, and show nice potential in controllable era duties.
