DPDDI: a Deep Predictor for Drug-Drug Interactions

Background— The treatment of complex diseases taking multiple drugs becomes popular. However, drug-drug interactions (DDIs) may give rise to the risk of unanticipated adverse effects and even unknown toxicity.As DDI detection in the wet lab is expensive and time-consuming, computational DDIs prediction based on machine learning becomes a promising approach due to its low cost and fast running.Generally, most of the existing computational approaches construct drug features from diverse drug properties, which are costly obtained and not available in many cases. Result— To address this issue, by organizing DDIs a network, we propose a novel predicting approach, which can without drug property.It consists of a feature extractor based on graph convolution network(GCN) as well as a predictor based on deep neural network (DNN). The formercharacterizes drugs in a graph embedding space, where each drugwasrepresented as a low-dimensional latent feature vector capturing the topological relationship to its neighborhood drugs by GCN. The latter concatenates latent feature vectors of any two drugsas the feature vector of the corresponding drug pairs and trains a DNN to predict potential interactions. In the experiments, we first demonstrate that our DNN-based predictor greatly outperforms the inner product-based predictor in the original GCN, and our network-derived latent feature greatly outperforms other features derived from chemical, biological or anatomicalproperties of drugs. Then, we indicate the over-optimistic prediction caused by down-sampling unlabeled drug pairs and validate the robustness of our approach to different datasets w.r.t. drug number, DDI number, and network sparsity. Moreover, the comparison with four state-of-the-art approaches using drug properties demonstrates the significant superiority of our approach under 5-fold cross-validation. Finally, a novel prediction validates its potentials in a real predicting scenario with finding 13 verified DDI out of the top 20 unlabeled candidates. Conclusion — We propose a simple but robust method DPDDI to predicting novel DDIs, which canwork without drug property. It can be expected that DPDDI can be helpful in other DDI-related scenarios, such as the detection of unexpected side effects, and the guidance of drug combination.