A Machine Learning Framework for Predicting Drug-drug Interactions

Abstract

Understanding drug-drug interaction is an essential step to reduce the risk of adverse drug events before clinical drug co-prescription. Existing methods commonly integrate multiple heterogeneous data sources to increase model performance but result in a high model complexity. To elucidate the molecular mechanisms behind drug-drug interactions and reserve rational biological interpretability is a major concern in computational modeling. In this study, we propose a simple representation of drug target profiles to depict drug pairs, based on which an l2-regularized logistic regression model is built to predict drug-drug interactions. In addition, we develop several statistical metrics to measure the communication intensity, interaction efficacy and action range between two drugs in the context of human protein-protein interaction networks and signaling pathways. Cross validation and independent test show that the simple feature representation via drug target profiles is effective to predict drug-drug interactions and outperforms the existing data integration methods. Statistical results show that two drugs easily interact when they target common genes, or their target genes communicate with each other via short paths in protein-protein interaction networks or through cross-talks between signaling pathways. The unravelled mechanisms provide biological insights into potential pharmacological risks of known drug-drug interactions and drug target genes.