High-Throughput Prediction of Drug-Drug Interactions from Molecular Structures Using Convolutional Neural Networks with Multi-Resolution Filters

Drug-drug interaction (DDI) occurs when two or more drugs influence each other’s performance which potentially alters their safety and efficacy. Predictions of these interactions are important in the medical field because they can lead to unexpected side effects, reduced effectiveness, increased blood levels, and other complications. There are four types of DDIs: mechanical, effect, advice, and no interaction. Identifying all possible DDIs in the current drug pool is challenging during the development of new drugs. Identifying potential interactions during clinical trials is time-consuming and resource-intensive, which contributes to the slowing pace of drug development. To address this problem, I proposed a convolutional neural network-based high-throughput system for predicting drug-drug interactions from molecular structures. The system takes the molecular structures of two drugs and predicts their possible interactions. It achieved an accuracy of 84.89% on a published dataset. Further analysis shows that the filter size of the convolutional neural network significantly impacts the system's accuracy.