Predicting drug-target interactions with deep neural networks in semi-supervised learning manner

Abstract

Identification of drug-target protein interactions (DTIs) is important in drug discovery fields. Diverse computational methods have been proposed to find out DTIs efficiently. The dataset generally used in training has only a small number of DTIs, so that many drug-target pairs have no interaction label. Most studies did not consider the mentioned characteristics of the DTI dataset and usually did not utilize the unlabeled data well. In this paper, we predict DTIs by using the features of drugs and targets. In the proposed prediction method, two main models proceed in sequence. The first model is the autoencoder model with multiple hidden layers. By only using the features of arbitrary drug-target pairs as the inputs, the autoencoder is trained without any label information in an unsupervised learning manner. The latent features from the autoencoder are used as the inputs of the second model. In the subsequent model, we train a deep neural network (DNN) classifier with the proposed semi-supervised learning method. Both unlabeled data and labeled data are utilized for training the classifier. The DNN classifier trained by the proposed unified model shows the best performance in all evaluation metrics compared to other supervised models utilizing only labeled data in training.