A Systematic Survey in Geometric Deep Learning for Structure-based Drug Design

Structure-based drug design (SBDD) utilizes the three-dimensional geometry of proteins to identify potential drug candidates. Traditional methods, grounded in physicochemical modeling and informed by domain expertise, are resource-intensive. Recent developments in geometric deep learning, focusing on the integration and processing of 3D geometric data, coupled with the availability of accurate protein 3D structure predictions from tools like AlphaFold, have greatly advanced the field of structure-based drug design. This paper systematically reviews the current state of geometric deep learning in SBDD. We first outline foundational tasks in SBDD, detail prevalent 3D protein representations, and highlight representative predictive and generative models. We then offer in-depth reviews of each key task, including binding site prediction, binding pose generation, \emph{de novo} molecule generation, linker design, and binding affinity prediction. We provide formal problem definitions and outline each task's representative methods, datasets, evaluation metrics, and performance benchmarks. Finally, we summarize the current challenges and future opportunities: current challenges in SBDD include oversimplified problem formulations, inadequate out-of-distribution generalization, a lack of reliable evaluation metrics and large-scale benchmarks, and the need for experimental verification and enhanced model understanding; opportunities include leveraging multimodal datasets, integrating domain knowledge, building comprehensive benchmarks, designing criteria based on clinical endpoints, and developing foundation models that broaden the range of design tasks. We also curate \url{https://github.com/zaixizhang/Awesome-SBDD}, reflecting ongoing contributions and new datasets in SBDD.