Faculty of Electrical Engineering and Computer Science, Ningbo University, Ningbo, 315211, China.
Engineering Laboratory of Advanced Energy Materials, Ningbo Institute of Materials Technology and Engineering, Chinese Academy of Sciences, Ningbo, 315201, China.
Comput Biol Med. 2023 Sep;164:107372. doi: 10.1016/j.compbiomed.2023.107372. Epub 2023 Aug 13.
Accurate prediction of drug-target affinity (DTA) plays a crucial role in drug discovery and development. Recently, deep learning methods have shown excellent predictive performance on randomly split public datasets. However, verifications are still required on this splitting method to reflect real-world problems in practical applications. And in a cold-start experimental setup, where drugs or proteins in the test set do not appear in the training set, the performance of deep learning models often significantly decreases. This indicates that improving the generalization ability of the models remains a challenge. To this end, in this study, we propose ColdDTA: using data augmentation and attention-based feature fusion to improve the generalization ability of predicting drug-target binding affinity. Specifically, ColdDTA generates new drug-target pairs by removing subgraphs of drugs. The attention-based feature fusion module is also used to better capture the drug-target interactions. We conduct cold-start experiments on three benchmark datasets, and the consistency index (CI) and mean square error (MSE) results on the Davis and KIBA datasets show that ColdDTA outperforms the five state-of-the-art baseline methods. Meanwhile, the results of area under the receiver operating characteristic (ROC-AUC) on the BindingDB dataset show that ColdDTA also has better performance on the classification task. Furthermore, visualizing the model weights allows for interpretable insights. Overall, ColdDTA can better solve the realistic DTA prediction problem. The code has been available to the public.
准确预测药物-靶标亲和力（DTA）在药物发现和开发中起着至关重要的作用。最近，深度学习方法在随机拆分的公共数据集上表现出了优异的预测性能。然而，这种拆分方法仍需要验证，以反映实际应用中的实际问题。并且在冷启动实验设置中，测试集中的药物或蛋白质不在训练集中，深度学习模型的性能通常会显著下降。这表明提高模型的泛化能力仍然是一个挑战。为此，在本研究中，我们提出了 ColdDTA：使用数据增强和基于注意力的特征融合来提高预测药物-靶标结合亲和力的泛化能力。具体来说，ColdDTA 通过删除药物的子图来生成新的药物-靶对。还使用基于注意力的特征融合模块来更好地捕获药物-靶相互作用。我们在三个基准数据集上进行了冷启动实验，Davis 和 KIBA 数据集上的一致性指数（CI）和均方误差（MSE）结果表明 ColdDTA 优于五种最先进的基线方法。同时，BindingDB 数据集上的接收器操作特征（ROC-AUC）的结果表明 ColdDTA 在分类任务上也具有更好的性能。此外，可视化模型权重可以提供可解释的见解。总体而言，ColdDTA 可以更好地解决现实世界的 DTA 预测问题。该代码已向公众开放。
