Advances in deep learning-based 3D molecular generative models

Abstract

<p indent="0mm">Molecular generation, aiming at designing novel molecules with desired properties such as biological activity and drug metabolism and pharmacokinetic (DMPK) properties in a low-cost and high-efficient manner, is a fundamental problem in drug discovery. Recently, deep generative models have been widely used in drug discovery, along with numerous model architectures and optimization strategies being explored, most of which are developed to generate one-dimensional or two-dimensional molecular structures. Inspired by the rapid development of deep learning in processing geometric graph data, 3D generative models for molecular generation have also been proposed, gaining attention for their advantages and potential on direct 3D molecular conformation generation and structure-based drug discovery. This survey has offered a systematic summarization of existing research achievements of the domestic and foreign researchers in recent years in the aspects of 3D molecular generation. According to the input of 3D molecular generation algorithms, it is divided into latent variable-based generation, 2D graph-based generation and 3D geometry-based generation. Then, according to the output of 3D molecular generation algorithms, it is divided into goal-directed generation and non-goal-directed generation. Furthermore, the performance of different generative models on main public datasets is summarized, which proves the advantages and disadvantages of the various models. Finally, some promising research directions are proposed.