Adverse Drug Reactions Prediction Using Multi-label Linear Discriminant Analysis and Multi-label Learning

Adverse Drug Reactions are responses to dangerous and undesirable drugs occurring at regular doses in humans for prophylaxis, diagnosis or treatment of disease, and modification of physiological functions. The identification of adverse drug reactions is crucial in the field of public health and pharmacology. Knowing the potential for early ADR can limit its effects on the patient's life as well as making the drug development pipeline stronger and more efficient. Several computational methods have been developed to predict the ADR potential from preclinical characteristics of compounds or filtering data. Because a drug can generally have several side effects, the ADR prediction can be formulated naturally as a multi-label learning problem. The high dimensions of the multi-label data make the reduction dimensions as a critical step in building an efficient and effective multi-label learning model. This research carried out feature dimension reduction using Multi-label Linear Discriminant Analysis and built multi-label learning models using three different multi-label learning algorithms, namely Multi-label k-Nearest Neighbor (MlkNN), Neural Networks (NN), and Deep Neural Networks (DNN). This study used the chemical structure of the drug sourced from PubChem and the SIDER database presenting information on the side effects and indications of marketed drugs (dataset 1). Moreover, the dataset in the study of Liu et al. (dataset 2) was used for comparison. The results of this study indicated that MLDA could reduce the dimensions of features and improve model performance. MLDA was able to reduce the feature dimension of dataset 1 from 881 features to 6 new features and dataset 2 from 2892 features to 6 new features. MLDA increased the AUC-ROC of model by 1.56% (MLKNN), 5.67% (NN), and 6.53% (DNN) respectively when using dataset 1. Then, when using dataset 2 MLDA increased the AUR-ROC of model by 3.88% (MLKNN), 6.67% (NN), and 6.90% (DNN) respectively.