Fig. 1
From: LOGICS: Learning optimal generative distribution for designing de novo chemical structures
An overview of the proposed LOGICS framework. In the pre-training phase, the prior generator is trained withâ~â1.6 million molecules from ChEMBL. In the fine-tuning phase, the agent generator is fine-tuned with the selected molecules from the generation of the agent itself and the experience memory, where the selection is performed with the three stages of tournament selections