Prediction of protein-protein interactions based on deep neural networks

In order to deal with the high false-positive to false-negative rate in experimental methods, a Deep Neural Network (DNN) is constructed based on several biology features. Protein features, including GO term semantic similarity, sequence similarity, essentiality and subcellular localization information, are integrated from diverse databases to form a fixed-length eigenvector. This vector contains a great deal of related information and can be used as the input of a classifier to predict protein interactions. Then the DNN which is data driven is constructed. It is used to automatically learn information from the input data and predict whether the unknown protein pairs interact or not. Dropout is used during the training phase to prevent co-adaption and improve its performance. The method achieves a prediction accuracy of 95.67% with 96.38% precision on the S. cerevisae dataset. Experimental results show that the extracted features are suitable for the prediction of PPIs, and many commonly used machine learning models can predict interaction effectively and efficiently based on this eigenvector. Moreover the DNN has good generalization capacity and shows high performance on various feature data. © 2019, Jilin University Press. All right reserved.