Application of generative adversarial networks (GAN) and reinforcement learning in drug classification

MolGAN is a generative adversarial network (GAN) model specially designed for molecular graph structures. It uses a graph convolutional network (GCN) as the discriminator and combines a reinforcement learning mechanism to optimize the generator. The model can efficiently process molecular graph data and generate new molecules with specific physicochemical properties. The experimental results show that MolGAN performs well in many aspects. First, MolGAN had significantly higher validity scores, uniqueness scores, and novelty scores than the baseline model in terms of quality of production, and the molecules produced were more chemically regular and diverse. Secondly, MolGAN outperformed the baseline model in terms of accuracy, precision, recall, F1 score, and AUC-ROC in classification performance, whether combined with SVM or DNN. MolGAN-generated molecules also performed well on bioactivity-related measures such as QED (score for quantitative drug similarity) and SA (synthetic accessibility score), showing better drug similarity and synthetic feasibility. In the diversity assessment, MolGAN generated molecules with lower average Tanimoto similarity and higher average Levenshtein distance, indicating more diversity.