Fig. 1

Schematic of this work including data sources (blue), scoring functions (orange), the deep generative model framework REINVENT [7] (grey). Main steps are (1) removing known DRD2 active molecules from the ZINC training data; (2) Training the Prior model on drug-like molecules from ZINC; (3) Initializing the Agents as a copy of the Prior; (4) Preparing the scoring functions to evaluate de novo molecules; (5) Iteratively training both Agents via reinforcement learning; and (6) evaluating the structure- and ligand-based approach with respect to different quantitative, chemical and structural aspects of the generated molecules