Drug repositioning framework using embedding drug-protein-disease similarities with graph convolution network and ensemble learning

Abstract

• Our proposed graph autoencoder uses a heterogeneous graph with multi-level attention structure and GCN to predict DTI. • The graph autoencoder has two components: a GAT convolution neural network with a multi-level attention structure as encoder. The GCN reconstructs the original graph using node embedding vectors as decoder. • To address data imbalance, we proposed a one-class classifier that predicts the most significant negative samples to balance the dataset. • Extensive tests with benchmark datasets are done to evaluate the performance of the suggested approach, which outperforms the closely related methods in terms of sensitivity, AUC and accuracy. The benefits of drug repositioning to the pharmaceutical industry have garnered significant attention in the field of drug development in recent years. Deep learning techniques have significantly improved drug repositioning by studying therapeutic drug profiles, diseases, and proteins. As the number of drugs increases, their targets and interactions generate imbalanced data, which may be undesirable as input to computational prediction model. The approach proposed in this paper uses a hierarchical network embedding technique and a graph autoencoder (GAE) scheme to solve this problem. The approach extracts embedding feature vectors of drugs and targets from a heterogeneous multi-source network to predict unknown drug-target interactions (DTIs). We employ a Meta-Path instance that has extensive drug and target characteristic data. The effectiveness of utilizing Meta-Path instance, the number of attention heads, and Graph Convolutional Network (GCN) and ensemble learning algorithm is analyzed on gold-standard datasets to evaluate the accuracy of the model and validity of the discovered DTI. The results achieved by our model using 10-fold cross-validation testing showed an improvement of 2.52 % in prediction accuracy, 4.2 % in recall, 3.94 % in AUC, and 3.6 % in F-score compared to state-of-the-art methods.