DecoyDB
收藏资源简介:
DecoyDB是一个大规模的、结构感知的数据集,专为蛋白质-配体复合物的自监督图对比学习而设计。该数据集由高分辨率真实3D复合物和多样化的诱饵结构组成,诱饵结构具有计算生成的结合姿势,范围从真实的(正对)到次优的(负对)。DecoyDB包含61104个真实3D复合物和5353307个诱饵,每个诱饵都标注了与原始姿势的均方根偏差(RMSD)。此外,还设计了一个定制的图对比学习算法,以基于DecoyDB对图神经网络进行预训练,并使用PDBbind的标签对模型进行微调。实验证明,使用DecoyDB预训练的模型在预测准确性、样本学习效率和泛化能力方面都有显著提升。
DecoyDB is a large-scale, structure-aware dataset specifically designed for self-supervised graph contrastive learning of protein-ligand complexes. This dataset comprises high-resolution experimentally determined 3D complexes and diverse decoy structures, where the decoys feature computationally generated binding poses ranging from native (positive pairs) to suboptimal (negative pairs). DecoyDB contains 61,104 authentic 3D complexes and 5,353,307 decoys, with each decoy annotated with the root-mean-square deviation (RMSD) relative to the native binding pose. Furthermore, a customized graph contrastive learning algorithm was developed to pre-train graph neural networks using DecoyDB, followed by fine-tuning the model with labels from PDBbind. Experimental results have shown that models pre-trained with DecoyDB exhibit substantial improvements in prediction accuracy, sample learning efficiency, and generalization capability.




