An entropy weight method to integrate big omics and mechanistically evaluate drug-induced liver injury

Abstract

Background & Aims: Drug-induced liver injury (DILI) accounts for more than half of acute liver failure cases in the United States and is a major health care issue to the public worldwide. As investigative toxicology is playing an evolving role in the pharmaceutical industry, mechanistic insights into drug hepatotoxicity can facilitate drug development and clinical medication. Methods: By integrating multi-source datasets including gene expression profiles of rat livers from open TG-GATEs and DrugMatrix, drug labels from FDA LTKB, and clinical reports from LiverTox, and with the employment of bioinformatic and computational tools, this study developed an approach to characterize and predict DILI based on the molecular understanding of the processes (toxicity pathways). Results: A panel of 11 pathways widely covering biological processes and stress responses was established using a training set of 6 positive and 1 negative DILI drugs from open TG-GATEs. An entropy weight method (EWM)-based model was developed to weight responsive genes within a pathway, and an interpretable machine learning model XGBoot-SHAP was trained to rank the importance of pathways to the panel activity. The panel activity was proven to differentiate between injured and non-injured sample points and characterize DILI manifestation using 6 training drugs. Next, the model was tested using additional 89 drugs (61 positives + 28 negatives), and a precision of 86% and higher can be achieved. Conclusions: This study provides a novel approach to mechanisms-driven prediction modeling, as well as big data integration for insights into pharmacology and other human biology areas. Yuan Jin and Yingqing Shou contributed equally to this manuscript.