GRU-based de novo molecular generation and combinatorial virtual screening of OX1R antagonists

Orexin-1 receptor (OX1R) has been proved to play an important role in the regulation of emotions, addiction, panic, or anxiety, and thus been a promising drug target for the treatments of drug addiction, anxiety, and depression, pain, and obesity. In this work, GRU-based deep neural network combined with transfer learning was successfully used to build a molecular generation model of OX1R antagonists by using 2,066,376 drug-like molecules from ChEMBL database and 11525 known OX1R antagonists. The results showed that the GRU-based generation model can accurately grasp the SMILES grammar of the drug-like molecules and tend to generate potential OX1R antagonists after transfer learning. Then, graph convolutional network (GCN) with multi-head attention mechanism followed by a cascade of traditional ligand, receptor, and rule-based virtual screening was performed to screen potent OX1R antagonists from the generated molecules, which results in 23 de novo potential OX1R antagonists with good drug-like and druggability properties. Overall, this paper integrates the advantages of traditional and data-driven drug design methods and provides important references for the lead compound discovery of OX1R antagonists. • The generative GRU model can generate potential OX1R antagonists efficiently. • Graph neural network with attention mechanism for building prediction model. • De novo antagonists with good drug-like and druggability properties.