Establishment of the image-guided antagonist discovery workflow. (A) Illustration of the virtual compound library generator process: known active molecules are fragmented into scaffolds and molecular fragments, which are then reassembled using a deep learning model (decorator) to generate diverse and novel molecules. These molecules form the virtual compound library for further screening. (B) The construction process of the screening model: the EP4 activity dataset was used to finetune the property prediction model, ImageMol, which had been pre-trained on a large-scale self-supervised dataset. This resulted in a predictor capable of determining whether a molecule is active. (C) Thumbnails of virtual filters: From top to bottom, the steps and corresponding changes in the number of compounds during the process of screening the final hit compounds from the large-scale generated dataset are shown. (D) Description of the process for further manual screening, chemical synthesis, and biological verification: the top 50 compounds identified by the screening model were selected for these subsequent steps.