Abstract
Drug-induced liver injury (DILI) is a major cause of the discontinuation of drug development, and a reliable method for the evaluation of DILI risk is highly needed. In this study, we have sought to develop a novel DILI evaluation method based on a read-across approach, in which the toxicity of an untested substance is predicted from the toxicity of its similar substances. From 287 drugs (213 DILI-positive and 74 DILI-negative) selected from the FDA's DILIrank data set, 10% were randomly selected as test drugs, and the rest were used as a reference set for similar drug selection. For the 287 drugs, HepG2 cell-based cytotoxicity assays and high-content analyses were performed, and inhibitory effects on human drug-metabolizing enzymes were determined. Using molecular descriptors calculated with alvaDesc, Euclidean distances between the drugs were calculated and neighbor (similar) drugs for each test drug were selected from the reference set. A test substance was judged as positive if the ratio of positive drugs in the neighbor drugs was higher than the positive rate of the entire data set. The sensitivity, specificity, and consistency rate were 0.714, 0.750, and 0.724, respectively. We further screened the neighbors for the 6 false- positive and 2 false-negative drugs by creating a decision tree using the in vitro assay data, and found that 6 substances were correctly predicted. These results suggest that read-across is useful for DILI prediction and that its accuracy can be improved by using in vitro assay results.
