A Knowledge Graph Embedding Approach for Polypharmacy Side Effects Prediction

Abstract

Predicting the side effects caused by drug combinations may facilitate the prescription of multiple medications in a clinical setting. So far, several prediction models of multidrug side effects based on knowledge graphs have been developed, showing good performance under constrained test conditions. However, these models usually focus on relationships between neighboring nodes of constituent drugs rather than whole nodes, and do not fully exploit the information about the occurrence of single drug side effects. The lack of learning the information on such relationships and single drug data may hinder improvement of performance. Moreover, compared with all possible drug combinations, the highly limited range of drug combinations used for model training prevents achieving high generalizability. To handle these problems, we propose a unified embedding-based prediction model using knowledge graph constructed with data of drug–protein and protein–protein interactions. Herein, single or multiple drugs or proteins are mapped into the same embedding space, allowing us to (1) jointly utilize side effect occurrence data associated with single drugs and multidrug combinations to train prediction models and (2) quantify connectivity strengths between drugs and other entities such as proteins. Due to these characteristics, it becomes also possible to utilize the quantified relationships between distant nodes, as well as neighboring nodes, of all possible multidrug combinations to regularize the models. Compared with existing methods, our model showed improved performance, especially in predicting the side effects of new combinations containing novel drugs that have no clinical information on polypharmacy effects. Furthermore, our unified embedding vectors have been shown to provide interpretability, albeit to a limited extent, for proteins highly associated with multidrug side effect.