Molecular generative Graph Neural Networks for Drug Discovery

Drug Discovery is a fundamental and ever-evolving field of research. The design new candidate molecules requires large amounts time money, computational methods are being increasingly employed to cut these costs. Machine learning ideal for the potential molecules, which naturally represented as graphs. Graph generation revolutionized by deep methods, molecular one its most promising applications. In this paper, we introduce sequential graph generator based on set neural network modules, call MG2N2. At each step, node or group nodes added graph, along with connections. modular architecture simplifies training procedure, also allowing an independent retraining single module. Sequentiality modularity make process interpretable. use Neural Networks maximizes information in input at generative consists subgraph produced during previous steps. Experiments unconditional QM9 Zinc datasets show that our model capable generalizing patterns seen phase, without overfitting. results indicate method competitive, outperforms challenging baselines generation.