Computational drug repositioning with attention walking

Abstract

Drug repositioning aims to identify new therapeutic indications for approved medications. Recently, the importance of computational drug repositioning has been highlighted because it can reduce the costs, development time, and risks compared to traditional drug discovery. Most approaches in this area use networks for systematic analysis. Inferring drug-disease associations is then defined as a link prediction problem in a heterogeneous network composed of drugs and diseases. In this article, we present a novel method of computational drug repositioning, named drug repositioning with attention walking (DRAW). DRAW proceeds as follows: first, a subgraph enclosing the target link for prediction is extracted. Second, a graph convolutional network captures the structural features of the labeled nodes in the subgraph. Third, the transition probabilities are computed using attention mechanisms and converted into random walk profiles. Finally, a multi-layer perceptron takes random walk profiles and predicts whether a target link exists. As an experiment, we constructed two heterogeneous networks with drug-drug similarities based on chemical structures and anatomical therapeutic chemical classification (ATC) codes. Using 10-fold cross-validation, DRAW achieved an area under the receiver operating characteristic (ROC) curve of 0.903 and outperformed state-of-the-art methods. Moreover, we demonstrated the results of case studies for selected drugs and diseases to further confirm the capability of DRAW to predict drug-disease associations.

Figure 1

The workflow of the proposed model, DRAW. A subgraph composed of the nodes of the target link and their neighbors is extracted from the original drug-disease heterogeneous network. Nodes in the subgraph are labeled by DRNL and fed into the GCN. Transition probabilities ${\mathit{P}}^{+}$ and ${\mathit{P}}^{-}$ are computed by the attention conducted on ${\mathit{G}}^{+}$ and ${\mathit{G}}^{-}$ with the features from the GCN. After a random walk is completed, all features are used as input of MLP, a binary graph classifier.

Figure 2

Accuracy comparison of the proposed method, DRAW, and the five state-of-the-art methods for drug-disease association prediction when network-ATC is used: (a) ROC curves on the drug-side, (b) Precision*-recall curves on the drug-side, (c) ROC curves on the disease-side, and (d) Precision*-recall curves on the disease-side.

Figure 3

Accuracy comparison of the proposed method, DRAW, and the five state-of-the-art methods for drug-disease association prediction when network-CS is used: (a) ROC curves on the drug-side, (b) Precision*-recall curves on the drug-side, (c) ROC curves on the disease-side, and (d) Precision*-recall curves on the disease-side.

Figure 4

The distributions of AUC values from 10 folds when the network density thresholds are 2%, 3%, 4%, and 5%. Drug-disease associations were predicted on the drug side with network-ATC by DRAW. The highest median AUC was achieved when the density threshold was 4%.