Yan Xiao-Ying, Zhang Shao-Wu
Key Laboratory of Information Fusion Technology of Ministry of Education, School of Automation, Northwestern Polytechnical University, Xi'an, 710072, China.
College of Computer Science, Xi'an Shiyou University, Xi'an, 710065, China.
Curr Protein Pept Sci. 2018;19(5):498-506. doi: 10.2174/1389203718666161108101118.
During the development process of new drugs, identification of the drug-target interactions wins primary concerns. However, the chemical or biological experiments bear the limitation in coverage as well as the huge cost of both time and money. Based on drug similarity and target similarity, chemogenomic methods can be able to predict potential drug-target interactions (DTIs) on a large scale and have no luxurious need about target structures or ligand entries.
In order to reflect the cases that the drugs having variant structures interact with common targets and the targets having dissimilar sequences interact with same drugs. In addition, though several other similarity metrics have been developed to predict DTIs, the combination of multiple similarity metrics (especially heterogeneous similarities) is too naïve to sufficiently explore the multiple similarities.
In this paper, based on Gene Ontology and pathway annotation, we introduce two novel target similarity metrics to address above issues. More importantly, we propose a more effective strategy via decision template to integrate multiple classifiers designed with multiple similarity metrics.
In the scenarios that predict existing targets for new drugs and predict approved drugs for new protein targets, the results on the DTI benchmark datasets show that our target similarity metrics are able to enhance the predictive accuracies in two scenarios. And the elaborate fusion strategy of multiple classifiers has better predictive power than the naïve combination of multiple similarity metrics.
Compared with other two state-of-the-art approaches on the four popular benchmark datasets of binary drug-target interactions, our method achieves the best results in terms of AUC and AUPR for predicting available targets for new drugs (S2), and predicting approved drugs for new protein targets (S3).These results demonstrate that our method can effectively predict the drug-target interactions. The software package can freely available at https://github.com/NwpuSY/DT_all.git for academic users.
在新药研发过程中，药物 - 靶点相互作用的识别备受关注。然而，化学或生物学实验在覆盖范围上存在局限性，且耗费大量时间和金钱。基于药物相似性和靶点相似性，化学基因组学方法能够大规模预测潜在的药物 - 靶点相互作用（DTIs），并且对靶点结构或配体条目没有过高要求。
为了反映具有不同结构的药物与共同靶点相互作用以及具有不同序列的靶点与相同药物相互作用的情况。此外，尽管已经开发了几种其他相似性度量来预测DTIs，但多个相似性度量（尤其是异质相似性）的组合过于简单，无法充分探索多种相似性。
在本文中，基于基因本体论和通路注释，我们引入了两种新的靶点相似性度量来解决上述问题。更重要的是，我们通过决策模板提出了一种更有效的策略，以整合使用多种相似性度量设计的多个分类器。
在预测新药的现有靶点以及预测新蛋白质靶点的获批药物的场景中，DTI基准数据集上的结果表明，我们的靶点相似性度量能够在这两种场景中提高预测准确性。并且多个分类器的精细融合策略比多个相似性度量的简单组合具有更好的预测能力。
在二元药物 - 靶点相互作用的四个流行基准数据集上，与其他两种最先进的方法相比，我们的方法在预测新药的可用靶点（S2）以及预测新蛋白质靶点的获批药物（S3）方面，在AUC和AUPR方面取得了最佳结果。这些结果表明我们的方法能够有效地预测药物 - 靶点相互作用。该软件包可供学术用户从https://github.com/NwpuSY/DT_all.git免费获取。
