Deep learning approaches for <italic>de novo</italic> drug design: an overview

Abstract

<p indent="0mm">The design and discovery of lead compounds is the most challenging and creative stage in drug development, and multiple attributes of candidate molecules need to be optimized in this process, such as structural novelty, biological activity, target selectivity, synthesability, druggability, and safety. Although the development and application of computer-aided drug design methods have substantially reduced the time and cost in the lead compound discovery stage, it has not yet been able to reverse the current situation of low success rate of drug development. In recent years, with the continuous development of deep learning (DL) technology, DL-based <italic>de novo</italic> drug design methods have brought new opportunities for the discovery of lead compounds. The DL frameworks used by these new drug design models include encoder-decoder, recurrent neural network, generative adversarial networks, and reinforcement learning. In this paper, the basic theory, the input molecular representation and the evaluation metrics of their methods are reviewed, and the application prospect of DL-based <italic>de novo</italic> drug design methods is also discussed.