How could a weighted drug-drug network help improve adverse drug reaction predictions? Machine learning reveals the importance of edge weights

Recent years have witnessed the booming data of drugs and their associated adverse drug reactions (ADRs), resulting in a comparatively high hospitalization rate worldwide. Therefore, a tremendous amount of research has been done to predict ADRs to keep the risks at a minimum. Due to the nature of the lab experiments being costly and time-consuming, researchers are looking forward to more extensive use of data mining and machine learning techniques in this field. This paper constructs a weighted drug-drug network based on an integration of various data sources. The network presents underlying relationships between drugs by creating connections between them according to their common ADRs. Then multiple node-level and graph-level network features are extracted from this network, e.g. weighted degree centrality, weighted PageRanks etc. By concatenating these features to the original drug features, it could be made possible to train and test seven classical machine learning algorithms, e.g. Logistic Regression, Random Forest, Support Vector Machine, etc. The experiments conclude that all the tested machine learning methods would benefit from adding those network measures, and the logistic regression (LR) model provides the highest mean AUROC score (0.821) across all ADRs in the experiment. Weighted degree centrality and weighted PageRanks are identified to be the most important network features in the LR classifier. These shreds of evidence strongly support that the network approach could be fundamental in future ADR prediction, and the network edge weights are important in the logistic regression model.