Drug-Induced Liver Injury (DILI) Classification and Its Application on Human DILI Risk Prediction

Abstract

Drug-induced liver injury (DILI) is one of the primary reasons for drugs being terminated in premarket studies or being withdrawn from the market after approval. Many new methodologies have been examined to improve DILI prediction, including high-throughput/high-content screening assays, in silico approaches and toxicogenomics, all of which rely on a truth set of drugs with well-defined DILI potential. However, defining a drug’s DILI risk is on varying interpretations, leading to differing classification schemes. Even when the same drug list is employed, variability in scheme affects predictive variables and models which lead to disparate risk prediction. Each model imbeds truth and knowledge in a different manner and context. An integrative approach melding models and variables should yield a system with enhanced prediction accuracy and better characterized mechanisms of action. Toward such an integrative predictor, we present four different classification schemes, i.e., an FDA labeling data-based approach (DILIrank dataset), a clinical evidence-based approach (LiverTox dataset), literature-based approaches (Greene and Xu datasets), and a registry-based approach (Suzuki dataset). Comparative analyses showed good general agreement between these approaches, with the most substantial difference observed between in silico models for drug-centric classification methods (i.e., drug-labeling and literature based approaches) versus clinical evidence-based methods (i.e., case reports and registry based approaches). The results suggest that substantial benefits can be obtained by consolidating various classification schemes to generate a larger dataset imbedding more diverse knowledge, and especially the new data streams from emerging technologies.