Nie Dou, Zhao Huifeng, Zhang Odin, Weng Gaoqi, Zhang Hui, Jin Jieyu, Lin Haitao, Huang Yufei, Liu Liwei, Li Dan, Hou Tingjun, Kang Yu
Innovation Institute for Artificial Intelligence in Medicine of Zhejiang University, College of Pharmaceutical Sciences, Zhejiang University, Hangzhou, 310058 Zhejiang, China.
Huawei Nanjing Research & Development Center, No. 101 Software Avenue, Yuhuatai District, Nanjing, 210012 Jiangsu, China.
J Chem Inf Model. 2025 Jan 13;65(1):173-186. doi: 10.1021/acs.jcim.4c02232. Epub 2024 Dec 16.
Three-dimensional (3D) molecular generation models employ deep neural networks to simultaneously generate both topological representation and molecular conformations. Due to their advantages in utilizing the structural and interaction information on targets, as well as their reduced reliance on existing bioactivity data, these models have attracted widespread attention. However, limited training and testing data sets and the unexpected biases inherent in single evaluation metrics pose a significant challenge in comparing these models in practical settings. In this work, we proposed Durian, an evaluation framework for structure-based 3D molecular generation that incorporates protein-ligand data with experimental affinity and a comprehensive array of physicochemical and geometric metrics. The benchmark tasks encompass assessing the capability of models to reproduce the property distribution of training sets, generate molecules with rational distributions of drug-related properties, and exhibit potential high affinity toward given targets. Binding affinities were evaluated using three independent docking methods (QuickVina2, Surflex and Gnina) with both "" and "" modes to reduce false positives arising from conformational searches or scoring functions. Specifically, we applied Durian to six 3D molecular generation methods: LiGAN, Pocket2Mol, DiffSBDD, SBDD, GraphBP, and SurfGen. While most methods demonstrated the ability to generate drug-like small molecules with reasonable physicochemical properties, they exhibited varying degrees of limitations in balancing novelty, structural rationality, and synthetic accessibility, thereby constraining their practical applications in drug discovery. Based on a total of 17 metrics, Durian highlights the importance of multiobjective optimization in 3D molecular generation methods. For instance, SurfGen and SBDD showed relatively comprehensive performance but could benefit from further improvements in molecular conformational rationality. Our evaluation framework is expected to provide meaningful guidance for the selection, optimization, and application of 3D generative models in practical drug design tasks.
三维（3D）分子生成模型利用深度神经网络同时生成拓扑表示和分子构象。由于它们在利用目标上的结构和相互作用信息方面具有优势，以及对现有生物活性数据的依赖减少，这些模型受到了广泛关注。然而，有限的训练和测试数据集以及单一评估指标中固有的意外偏差，在实际环境中比较这些模型时构成了重大挑战。在这项工作中，我们提出了Durian，这是一个基于结构的3D分子生成评估框架，它将蛋白质 - 配体数据与实验亲和力以及一系列全面的物理化学和几何指标相结合。基准任务包括评估模型再现训练集性质分布的能力、生成具有合理药物相关性质分布的分子以及对给定目标表现出潜在高亲和力的能力。使用三种独立的对接方法（QuickVina2、Surflex和Gnina）在“ ”和“ ”模式下评估结合亲和力，以减少由构象搜索或评分函数产生的假阳性。具体而言，我们将Durian应用于六种3D分子生成方法：LiGAN、Pocket2Mol、DiffSBDD、SBDD、GraphBP和SurfGen。虽然大多数方法都展示了生成具有合理物理化学性质的类药物小分子的能力，但它们在平衡新颖性、结构合理性和合成可及性方面表现出不同程度的局限性，从而限制了它们在药物发现中的实际应用。基于总共17个指标，Durian强调了多目标优化在3D分子生成方法中的重要性。例如，SurfGen和SBDD表现出相对全面的性能，但在分子构象合理性方面仍可进一步改进。我们的评估框架有望为3D生成模型在实际药物设计任务中的选择、优化和应用提供有意义的指导。
