Abstract
Structure-based drug design (SBDD), which utilizes the three-dimensionalgeometry of proteins to identify potential drug candidates, is becomingincreasingly vital in drug discovery. However, traditional methods based onphysiochemical modeling and experts' domain knowledge are time-consuming andlaborious. The recent advancements in geometric deep learning, which integratesand processes 3D geometric data, coupled with the availability of accurateprotein 3D structure predictions from tools like AlphaFold, have significantlypropelled progress in structure-based drug design. In this paper, wesystematically review the recent progress of geometric deep learning forstructure-based drug design. We start with a brief discussion of the mainstreamtasks in structure-based drug design, commonly used 3D protein representationsand representative predictive/generative models. Then we delve into detailedreviews for each task (binding site prediction, binding pose generation,\emph{de novo} molecule generation, linker design, and binding affinityprediction), including the problem setup, representative methods, datasets, andevaluation metrics. Finally, we conclude this survey with the currentchallenges and highlight potential opportunities of geometric deep learning forstructure-based drug design.We curate a GitHub repo containing the relatedpapers \url{https://github.com/zaixizhang/Awesome-SBDD}.