Department of Computer Science and Engineering, School of Information Science and Engineering, Yunnan University, Kunming, 650504, Yunnan, China.
Department of Computer Science and Engineering, School of Information Science and Engineering, Yunnan University, Kunming, 650504, Yunnan, China; The Key Lab of Intelligent Systems and Computing of Yunnan Province, Yunnan University, Kunming, Yunnan, China.
Comput Biol Chem. 2023 Dec;107:107968. doi: 10.1016/j.compbiolchem.2023.107968. Epub 2023 Oct 7.
Predicting drug-target interaction (DTI) is a crucial step in the process of drug repurposing and new drug development. Although the attention mechanism has been widely used to capture the interactions between drugs and targets, it mainly uses the Simplified Molecular Input Line Entry System (SMILES) and two-dimensional (2D) molecular graph features of drugs. In this paper, we propose a neural network model called MdDTI for DTI prediction. The model searches for binding sites that may interact with the target from the multiple dimensions of drug structure, namely the 2D substructures and the three-dimensional (3D) spatial structure. For the 2D substructures, we have developed a novel substructure decomposition strategy based on drug molecular graphs and compared its performance with the SMILES-based decomposition method. For the 3D spatial structure of drugs, we constructed spatial feature representation matrices for drugs based on the Cartesian coordinates of heavy atoms (without hydrogen atoms) in each drug. Finally, to ensure the search results of the model are consistent across multiple dimensions, we construct a consistency loss function. We evaluate MdDTI on four drug-target interaction datasets and three independent compound-protein affinity test sets. The results indicate that our model surpasses a series of state-of-the-art models. Case studies demonstrate that our model is capable of capturing the potential binding regions between drugs and targets, and it shows efficacy in drug repurposing. Our code is available at https://github.com/lhhu1999/MdDTI.
预测药物-靶标相互作用（DTI）是药物再利用和新药开发过程中的关键步骤。尽管注意力机制已被广泛用于捕捉药物和靶标之间的相互作用，但它主要使用简化分子输入行系统（SMILES）和药物的二维（2D）分子图特征。在本文中，我们提出了一种名为 MdDTI 的神经网络模型，用于 DTI 预测。该模型从药物结构的多个维度（即 2D 子结构和三维（3D）空间结构）搜索可能与靶标相互作用的结合位点。对于 2D 子结构，我们基于药物分子图开发了一种新的子结构分解策略，并将其性能与基于 SMILES 的分解方法进行了比较。对于药物的 3D 空间结构，我们基于每个药物中重原子（不含氢原子）的笛卡尔坐标构建了药物的空间特征表示矩阵。最后，为了确保模型的搜索结果在多个维度上保持一致，我们构建了一个一致性损失函数。我们在四个药物-靶标相互作用数据集和三个独立的化合物-蛋白亲和力测试集上评估了 MdDTI。结果表明，我们的模型优于一系列最先进的模型。案例研究表明，我们的模型能够捕捉药物和靶标之间的潜在结合区域，并且在药物再利用方面表现出了疗效。我们的代码可在 https://github.com/lhhu1999/MdDTI 上获得。
