Context-Aware Multilevel Classification of Semantic Relations in Drug-Adverse Drug Reaction (ADR) Networks-Predicting Drug-Induced Liver Injury (DILI) as a Case Study.

Adverse drug reactions (ADRs) are a major concern for public health and patient safety and a burden in drug discovery, with drug-induced liver injury (DILI) as a prototypical case. Next to computational models such as deep learning, knowledge graph models can help capture biological systems and pathways underlying toxic effects. We present a hierarchical, context-aware knowledge graph model to detect ADRs, with DILI as a case study. The model builds on a commercial knowledge graph integrating over 200 life-science databases and millions of publications and patents. It combines translation- and path-based methods, where two-hop paths link drugs and ADRs via an intermediate concept (e.g., genes, proteins, pathways). These paths are combined in path bundles as the graph's basic units. Evaluation used the FDA DILIrank data set of DILI-positive and DILI-negative drugs, mapped onto the graph. Performance was evaluated with classification metrics, including the area under the receiver operating characteristic curve (AUC) and Matthews correlation coefficient (MCC). The output probabilities of the path bundles generated by base classifiers formed the input for multilevel statistical modeling. Multilevel analysis was used to predict DILI risk of the drugs while accounting for the dependencies of the path bundles on the ADR contexts and on the different relations in the graph. The four-level model performed best (AUC = 0.850; MCC = 0.639). Average prediction probabilities showed good separation (0.641 for DILI-positive drugs and 0.402 for DILI-negative drugs). Additional results included the correct classification of 29 of the 37 wrongly classified drugs by a QSAR study using molecular descriptors. The added value of a cascaded QSAR and knowledge graph approach was shown on a 130-compound subset of the DeepDILI test set, where the combined model rescued additional DILI-positive drugs missed by DeepDILI alone at the expense of a limited number of false positives. The study model and data are available through https://github.com/mi-erasmusmc/multilevel_dili.