Fig. 3

The overall experimental procedure. a The structural and physicochemical descriptors of the dataset were calculated. Both features were used as the input, and the DILI label as the output. The RF, LGBM, LR, and NN with attention were used. To interpret the models, b identifying important features of the overall DILI predictions using the permutation feature importance and c highlighting specific molecular substructures were conducted by analyzing the attention weight