Fig. 1
From: Prediction and mechanistic analysis of drug-induced liver injury (DILI) based on chemical structure
DILI label prediction performance (balanced accuracy) of RF and SVM models trained using ECFP4 descriptors. Models were trained using the datasets described in Table 1. Performance is stable between the 5-fold LOCO-CV and external test set, but a distinct drop in predictive accuracy is observed when predicting the FDA validation set. Hence, despite demonstrating a capability to generalize to new compounds (not seen during training) in the external test set, models lacked the capability to generalize to the new compounds in the FDA validation set