Wang Guishen, Zhang Hangchen, Shao Mengting, Sun Shisen, Cao Chen
College of Computer Science and Engineering, Changchun University of Technology, North Yuanda Street No. 3000, Jilin 130012, China.
Key Laboratory for Bio-Electromagnetic Environment and Advanced Medical Theranostics, School of Biomedical Engineering and Informatics, Nanjing Medical University, Longmian Avenue No. 101, Jiangsu 211166, China.
J Chem Inf Model. 2025 Feb 10;65(3):1615-1630. doi: 10.1021/acs.jcim.4c01528. Epub 2025 Jan 20.
Predicting drug-target binding affinity (DTA) is a crucial task in drug discovery research. Recent studies have demonstrated that pocket features and interactions between targets and drugs significantly improve the understanding of DTA. However, challenges remain, particularly in the detailed consideration of both global and local information and the further modeling of pocket features. In this paper, we propose a novel multimodal deep learning model named MMPD-DTA for predicting drug-target binding affinity to address these challenges. The MMPD-DTA model integrates graph and sequence modalities of targets, pockets, and drugs to capture both global and local target and drug information. The model introduces a novel pocket-drug graph (PD graph) that simultaneously models atomic interactions within the target, within the drug, and between the target and drug. We employ GraphSAGE for graph representation learning from the PD graph, complemented by sequence representation learning via transformers for the target sequence and graph representation learning via a graph isomorphism network for the drug molecular graph. These multimodal representations are then concatenated, and a multilayer perceptron generates the final binding affinity predictions. Experimental results on three real-world test sets demonstrate that the MMPD-DTA model outperforms baseline methods. Ablation studies further confirm the effectiveness of each module within the MMPD-DTA model. Our code is available at https://github.com/zhc-moushang/MMPD-DTA.
预测药物-靶点结合亲和力（DTA）是药物发现研究中的一项关键任务。最近的研究表明，靶点特征以及靶点与药物之间的相互作用显著提升了对DTA的理解。然而，挑战依然存在，尤其是在全面考虑全局和局部信息以及进一步对口袋特征进行建模方面。在本文中，我们提出了一种名为MMPD-DTA的新型多模态深度学习模型，用于预测药物-靶点结合亲和力，以应对这些挑战。MMPD-DTA模型整合了靶点、口袋和药物的图结构与序列模态，以捕捉靶点和药物的全局和局部信息。该模型引入了一种新型的口袋-药物图（PD图），它同时对靶点内部、药物内部以及靶点与药物之间的原子相互作用进行建模。我们使用GraphSAGE从PD图中进行图表示学习，并通过针对靶点序列的transformer进行序列表示学习，以及通过图同构网络对药物分子图进行图表示学习来加以补充。然后将这些多模态表示连接起来，由多层感知器生成最终的结合亲和力预测结果。在三个真实世界测试集上的实验结果表明，MMPD-DTA模型优于基线方法。消融研究进一步证实了MMPD-DTA模型中每个模块的有效性。我们的代码可在https://github.com/zhc-moushang/MMPD-DTA获取。
