Chapter 10: Compound Design Using Generative Neural Networks

In recent years, deep learning has substantially impacted fields such as image analysis or natural language processing. Inspired by these successes, computational chemists are increasingly adopting generative modeling techniques to produce new molecules and predict their properties. In this chapter, we describe the use of recurrent neural networks, reinforcement learning, and autoencoder networks in compound design. First, we introduce the foundation of deep learning and the most commonly used molecular representations. Next, we explain how generative models can be applied to create novel compounds. Then, we discuss recent advances in molecular property prediction, which is used to guide generative models during compound design. Accordingly, accurate prediction of molecular properties is an integral part of compound generation via deep learning.