Abstract: The prediction of compound-protein interaction (CPI) is a critical technological tool for discovering lead compounds and drug repurposing during the process of drug development.In recent years, deep learning has been widely used in CPI research, which has accelerated the development of CPI prediction in drug discovery.This review focuses on feature-based CPI prediction models.First, we described the datasets, as well as typical feature representation methods commonly used for compounds and proteins in CPI prediction.Based on the critical problems in modeling, we discussed models for CPI prediction from two perspectives: multimodal features and attention mechanisms.Then, the performance of 12 selected models was evaluated on 3 benchmark datasets for both classification and regression tasks.Finally, the review summarizes the existing challenges in this field and prospects for future directions.We believe that this investigation will provide some reference and insight for further research on CPI prediction.