A general prediction model of compound-protein interaction based on deep learning

Compound-protein interaction (CPI) is an important basis for drug discovery and repurposing. Accurate identification of CPI can significantly reduce the time and cost of discovering candidate compounds and advance drug research and development. In recent years, in silico methods have been applied to predict potential CPI, but the accuracy and the generalization capability of prediction models is still a standing challenge due to the structural diversity of compounds and targets, and the lack of large-scale known interactions for model learning. To address these challenges, we developed an efficient, accurate and highly applicable computational model for CPI prediction by combining the constructed large-scale bioactivity benchmark dataset with deep learning (DL) algorithm. Our CPI prediction model signi-    ficantly outperformed other machine learning models, achieving an area under the Receiver Operating Characteristic curve (AUROC) of 0.98, an area under the precision-recall curve (AUPR) of 0.98, and an accuracy (ACC) of 93.31% for the test set. Furthermore, the generalization capability and applicability of the CPI prediction model were investigated using external databases. The results indicated that our model was capable of making accurate prediction for novel CPI data outside the training dataset and evaluating the reliability of the prediction results, showing a good applicability potential. Our CPI prediction model can serve as a useful tool for accurately identifying potential CPI for wide range of proteins, which is expected to facilitate drug research and repurposing.