Fig. 1

Overall data collection process. a First, we collected the public five datasets. b DILI positives and DILI negatives were selected by filtering drugs and compounds clearly related to DILI. c After integrating the datasets, duplicates and unclear structural formats were removed. Additionally, when the class label of compounds had a conflict, the DILIrank or NCTR data label was assigned. d NCTR, Greene, Xu, and Liew datasets were used as the training set and the DILIrank dataset was used as the test set