Predicting Adverse Drug Reactions from Drug Functions by Binary Relevance Multi-label Classification and MLSMOTE

Adverse Drug Reaction (ADR) prediction is one of the important tasks in drug discovery. It helps in enhancing drug safety and reducing drug discovery costs and time. Most of the existing works have focused on ADR prediction using chemical and biological properties of drugs. However, the capability of drug functions in ADR prediction has not been explored yet. ADR prediction is a multi-label classification problem and it faces the issue of class imbalance. In the present work, a methodology has been proposed for predicting ADR from drug functions. It employs the binary relevance method along with five base classifiers namely DT, ETC, KNN, MLPNN, and RF for performing multi-label classification and MLSMOTE for addressing the issue of class imbalance. The data of drug functions and ADR has been extracted respectively from SIDER and PubChem databases and then drug functions are mapped to ADR based on drug ID. After mapping drug function with the ADR, the resulted dataset comprises 670 drugs described by their functions and 6123 ADR. The proposed methodology has been applied on this dataset. The performance of the proposed methodology has been found promising in terms of accuracy, hamming loss, precision, recall, f1 score and ROC-AUC.