MAARDTI: a multi-perspective attention aggregation model for the prediction of drug–target interactions

Accurate prediction of drug–target interactions (DTIs) is indispensable for discovering novel drugs and repositioning existing ones. Recently, numerous methods based on deep learning have made promising progress in DTI predictions. These methods often utilize a single attention mechanism, which limits their ability to capture the complex features of both drugs and proteins. As a result, feature representation can be incomplete, training can become more complex and prone to overfitting. These together can impair the generalizability of the model. To address these problems, we propose an end-to-end neural network drug–target interaction approach called Multi-perspective Attention AggRegating (MAARDTI). Here, a multi-perspective attention mechanism is introduced that combines channel attention and spatial attention to capture a more comprehensive feature representation. The dual-context refocusing module is used to enhance the attention representation capability and improve the generalizability of the model. Experiments show that our proposed model outperforms ten state-of-the-art methods in three public datasets, achieving AUC values of 0.8975, 0.9248, and 0.9330 in DrugBank, Davis and KIBA, respectively. In the cold-splitting test with novel targets, drugs, and their bindings, MAARDTI performs on par with some methods for cold drug predictions. It outperforms in predicting unseen targets and bindings, underscoring the effectiveness of the novel multi-perspective attention mechanism in challenging scenarios. Hence, MAARDTI has the potential to serve as an effective tool for rapid identification of novel DTIs in drug research.