Drug-Target Interaction Prediction With Non-Random Missing Theory

In recent years,drug-target interaction(DTI)prediction is a very hot research topic.The predictions can be used to discover new drugs as well as find new targets for existing drugs.The majority of existing machine learning-based methods treat DTI prediction as a binary classification task.However,due to the large number of unknown labels samples,this approach will reduce the accuracy of the model.Our assumption in this work is that labels are not missing at random,because researchers will use their domain expertise to filter DTIs with a high possibility to be positive and prioritize validations for these DTIs in vivo,and those that they think may be negative samples will not be studied,making the labels of these negative sample pairs more likely to be missing.Based on the not missing at random data theory introduced above,this paper proposes a probabilistic model for the prediction of drug target interaction.The model in this paper models the labels and responses of drug-target interaction.At the same time,in order to reduce the variance of the model and improve the prediction accuracy of the model to adapt to the highly unbalanced drug-target data,this paper further proposes an ensemble method.The method in this paper is different from the traditional Boosting and Bagging ensemble methods.The proposed model was validated on the latest drug-target dataset.The experimental results show that the model of this paper is significantly better than other drug-target interaction prediction methods.At the same time,for different features,our model can improve the performance steadily.The ensemble method proposed in this paper can also significantly improve the model performance.Finally,this paper also use FNML model on the recommendation system and achieve good results,showing the good adaptability of the model.