Retrieval of a new starting active compound with a novel scaffold during early drug development is an important but challenging task. Here, the authors propose a generative deep learning model and by applying this model they discover a potent and highly selective RIPK1 inhibitor with a previously unreported scaffold.