QLDTI: A Novel Reinforcement Learning-based Prediction Model for Drug-Target Interaction

Background: Predicting drug-target interaction (DTI) plays a crucial role in drug research and development. More and more researchers pay attention to the problem of developing more powerful prediction methods. Traditional DTI prediction methods are basically realized by biochemical experiments, which are time-consuming, risky, and costly. Nowadays, DTI prediction is often solved by using a single information source and a single model, or by combining some models, but the prediction results are still not accurate enough. Objective: The study aimed to utilize existing data and machine learning models to integrate heterogeneous data sources and different models, further improving the accuracy of DTI prediction. Methods: This paper has proposed a novel prediction method based on reinforcement learning, called QLDTI (predicting drug-target interaction based on Q-learning), which can be mainly divided into two parts: data fusion and model fusion. Firstly, it fuses the drug and target similarity matrices calculated by different calculation methods through Q-learning. Secondly, the new similarity matrices are inputted into five models, NRLMF, CMF, BLM-NII, NetLapRLS, and WNN-GIP, for further training. Then, all sub-model weights are continuously optimized again by Q-learning, which can be used to linearly weight all sub-model prediction results to output the final prediction result. Results: QLDTI achieved AUC accuracy of 99.04%, 99.12%, 98.28%, and 98.35% on E, NR, IC, and GPCR datasets, respectively. Compared to the existing five models NRLMF, CMF, BLM-NII, NetLapRLS, and WNN-GIP, the QLDTI method has achieved better results on four benchmark datasets of E, NR, IC, and GPCR. Conclusion: Data fusion and model fusion have been proven effective for DTI prediction, further improving the prediction accuracy of DTI.