Wankhade Nitin, Sharma Anju, Wani Mushtaq Ahmad, Banerjee Aritra, Garg Prabha
Department of Pharmacoinformatics, National Institute of Pharmaceutical Education and Research, S. A. S. Nagar, Punjab 160 062, India.
ACS Med Chem Lett. 2024 Oct 2;15(11):1907-1917. doi: 10.1021/acsmedchemlett.4c00358. eCollection 2024 Nov 14.
Diabetes mellitus (DM) is a global health concern, and dipeptidyl peptidase-4 (DPP-4) is a key therapeutic target. The study used three machine learning and deep learning models to predict potential DPP-4 inhibitors using a curated data set of 6,750 compounds. The models included support vector machine (SVM), random forest (RF), naive Bayes (NB), and multitask deep neural network (MTDNN). The MTDNN model demonstrated strong predictive performance, achieving 98.62% train accuracy and 98.42% test accuracy for predicting DPP-4 inhibitors and a correlation coefficient of 0.979 for training and 0.977 for the test data set, with low training and test errors while predicting corresponding IC values. The MTDNN model predicted potential inhibitors using an external data set of FDA-approved drugs, identifying 100 compounds. Among these, five compounds stood out with promising molecular docking and dynamic profiles, suggesting their potential as repurposed drugs for targeting DPP-4 and offering hope for the future of diabetes treatment.
糖尿病（DM）是一个全球性的健康问题，而二肽基肽酶-4（DPP-4）是一个关键的治疗靶点。该研究使用了三种机器学习和深度学习模型，利用一个由6750种化合物组成的精选数据集来预测潜在的DPP-4抑制剂。这些模型包括支持向量机（SVM）、随机森林（RF）、朴素贝叶斯（NB）和多任务深度神经网络（MTDNN）。MTDNN模型表现出强大的预测性能，在预测DPP-4抑制剂时，训练准确率达到98.62%，测试准确率达到98.42%，训练数据集的相关系数为0.979，测试数据集的相关系数为0.977，在预测相应的IC值时训练和测试误差较低。MTDNN模型使用FDA批准药物的外部数据集预测潜在抑制剂，识别出100种化合物。其中，有五种化合物在分子对接和动力学方面表现出色，表明它们作为靶向DPP-4的重新利用药物的潜力，并为糖尿病治疗的未来带来了希望。
