A Simple Way to Incorporate Target Structural Information in Molecular Generative Models

Deep learning generative models arenow being appliedin variousfields including drug discovery. In this work, we propose a novelapproach to include target 3D structural information in moleculargenerative models for structure-based drug design. The method combinesa message-passing neural network model that predicts docking scoreswith a generative neural network model as its reward function to navigatethe chemical space searching for molecules that bind favorably witha specific target. A key feature of the method is the constructionof target-specific molecular sets for training, designed to overcomepotential transferability issues of surrogate docking models througha two-round training process. Consequently, this enables accurateguided exploration of the chemical space without reliance on the collectionof prior knowledge about active and inactive compounds for the specifictarget. Tests on eight target proteins showed a 100-fold increasein hit generation compared to conventional docking calculations andthe ability to generate molecules similar to approved drugs or knownactive ligands for specific targets without prior knowledge. Thismethod provides a general and highly efficient solution for structure-basedmolecular generation.