DeepERA: deep learning enables comprehensive identification of drug-target interactions via embedding of heterogeneous data

Drug-target interaction prediction is a crucial step in drug development, discovery, and repurposing. Due to the tremendous combinatorial search space of all drug-protein pairs, machine learning algorithms have been utilized to facilitate the identification of novel drug-target interactions. Deep learning, known as a powerful learning system, has recently shown superior performance to traditional machine learning in many biological and biomedical areas. In this paper, we proposed an end-to-end deep learning model, DeepERA, to identify drug-target interactions based on heterogeneous data. This model assembles three independent feature embedding modules (intrinsic embedding, relational embedding, and annotation embedding) which each represent different attributes of the dataset and jointly contribute to the comprehensive predictions. This is the first work that, to our knowledge, applied deep learning models to learn each intrinsic features, relational features, and annotation features and combine them to predict drug-protein interactions. Our results showed that DeepERA outperformed other deep learning approaches proposed recently. The studies of individual embedding modules explained the dominance of DeepERA and confirmed the effects of the "guilt by associations" assumption on the performance of the prediction model. Using our DeepERA framework, we identified 45,603 novel drug-protein interactions for the whole human proteome, including 356 drug-protein interactions for the human proteins targeted by SARS-CoV-2 viral proteins. We also performed computational docking for the selected interactions and conducted a two-way statistical test to "normalize" the docking scores of different proteins/drugs to support our predictions.