Chapter 13. Molecular De Novo Design Through Deep Generative Models

Abstract

Molecular de novo design through deep learning has become very popular since 2017. The methods have entirely transformed the possibilities to sample the chemical space to identify novel compounds with desirable properties. This chapter describes the key deep learning architectures for molecular de novo design, such as recurrent neural networks, generative adversarial networks, or variational autoencoders. Both string- and graph-based methods are discussed. Furthermore, methods to bias the generation to the desirable chemical space such as transfer learning and reinforcement learning are reviewed with a large emphasis on benchmarking methodologies to compare the performance of the different deep learning architectures. Several benchmarking methodologies have been developed during recent years comparing how the different deep learning architectures sample the chemical space. An important part is benchmarking how well recurrent neural networks cover the chemical space. In particular, we show that for a fragment-like molecular database, recurrent neural networks cover most of the chemical space. This chapter describes one of the most exciting developments in drug design that makes the dream come true, where we are able to sample the whole chemical space of interest without exhaustive enumeration.