Implementation of deep learning in drug design

The field of deep learning has witnessed dramatic and rapid progress in the past several years, largely driven by the availability of massive datasets and increased computational power. Although promising advances have been made, the implementation of deep learning in drug design is still in the vigorously developing stage. Here we summarize the frontiers and emerging applications of deep learning in drug design. We provide the background of deep learning and the architecture of several important deep neural networks for drug design and discovery, including convolutional neural networks, recurrent neural networks, transformer, deep auto‐encoder, generative adversarial networks, graph neural networks, and flow‐based neural networks. We give some examples of tasks for which these different deep neural networks are suitable. The successful employment of deep learning in drug design and discussing its further opportunities are reviewed, such as molecular representation, drug‐likeness prediction, de novo molecular generation, ligand‐protein interaction prediction, reaction, and retrosynthetic prediction. Current challenges and further possibilities are proposed to help accelerate the progress of the field, such as model interpretation, the requirement for large quantities of training datasets, hyperparameters tuning, and open source. We review the architecture of important deep neural networks for drug design and discovery, including convolutional neural networks, recurrent neural networks, transformers, deep auto‐encoder, generative adversarial networks, graph neural networks, and flow‐based neural networks. We discuss the successful employment and further opportunities of deep learning in molecular representation, drug‐likeness prediction, de novo molecular generation, ligand‐protein interaction prediction, reaction, and retrosynthetic prediction.