Prediction of Drug-Drug Interactions Based on Multi-layer Feature Selection and Data Balance

Drug-drug interactions (DDIs) occur when two drugs react with each other, which may cause unexpected side effects and even death of the patient. Methods that use adverse event reports to predict unexpected DDIs are limited by two critical yet challenging issues. One is the difficulty of selecting discriminative features from numerous redundant and irrelevant adverse events for modeling. The other is the data imbalance, i.e., the drug pairs causing adverse effects are far less than those not causing adverse effects, which leads to poor accuracy of DDIs detection. We propose a multi-layer feature selection method to select discriminative adverse events and apply an over-sampling technique to make the data balanced. The experimental results show that the validation accuracy of positive DDIs on the Canada Vigilance Adverse Reaction Online Database increases to two times, and 110 DDIs are identified on the drug interactions checker of Drugs.com in USA.