Predicting drug−disease associations via sigmoid kernel-based convolutional neural networks

• Published in: Journal of translational medicine Vol. 17; no. 1; pp. 382 - 11
• Main Authors: Jiang, Han-Jing, You, Zhu-Hong, Huang, Yu-An
• Format: Journal Article
• Language: English
• Published: England BioMed Central Ltd 20.11.2019; BioMed Central; BMC
• Subjects: Analysis; Artificial neural networks; Asthma; Bioinformatics; Cable television broadcasting industry; Case studies; Computational biology; Computer applications; Convolutional Neural Networks; Data mining; Datasets; Deep learning; Disease; Drug development; Drugs; Gene expression; Medical research; Methods; Natural language processing; Neural networks; Performance evaluation; Prediction models; Random forest; Recommender systems; Semantics; Sigmoid kernel; United States; Random forest; Sigmoid kernel; Convolutional Neural Networks
• DOI: 10.1186/s12967-019-2127-5
• ISSN: 1479-5876

In the process of drug development, computational drug repositioning is effective and resource-saving with regards to its important functions on identifying new drug-disease associations. Recent years have witnessed a great progression in the field of data mining with the advent of deep learning. An increasing number of deep learning-based techniques have been proposed to develop computational tools in bioinformatics. Along this promising direction, we here propose a drug repositioning computational method combining the techniques of Sigmoid Kernel and Convolutional Neural Network (SKCNN) which is able to learn new features effectively representing drug-disease associations via its hidden layers. Specifically, we first construct similarity metric of drugs using drug sigmoid similarity and drug structural similarity, and that of disease using disease sigmoid similarity and disease semantic similarity. Based on the combined similarities of drugs and diseases, we then use SKCNN to learn hidden representations for each drug-disease pair whose labels are finally predicted by a classifier based on random forest. A series of experiments were implemented for performance evaluation and their results show that the proposed SKCNN improves the prediction accuracy compared with other state-of-the-art approaches. Case studies of two selected disease are also conducted through which we prove the superior performance of our method in terms of the actual discovery of potential drug indications. The aim of this study was to establish an effective predictive model for finding new drug-disease associations. These experimental results show that SKCNN can effectively predict the association between drugs and diseases.