Fig. 1

The workflow of the training process of our deep learning-based molecule generator DrugEx2 utilizing reinforcement learning. After the generator has been pre-trained/fine-tuned, (1) a batch of SMILES are generated by sampling tokens step by step based on the probability calculated by the generator; (2) These valid SMILES are parsed to be molecules and encoded into descriptors to get the predicted pXs with predictors; (3) The predicted pXs are transformed into a single value as the reward for each molecule based on Pareto optimization; (4) These SMILES sequences and their rewards are sent back to the generator for training with policy gradient methods. These four steps constitute the training loop of reinforcement learning