Skip to main content

Table 3 Comparison of performance indicators to several DILI classification models reported in the literature

From: Prediction and mechanistic analysis of drug-induced liver injury (DILI) based on chemical structure

Model algorithm Number of Compounds CV scheme CV Balanced accuracy  CV Sensitivity External test set Balanced accuracy External test set Sensitivity References
RF 996 (541+/455-) 10-fold, random splits 0.645 0.680 0.588 0.536 Kotsampasakou et al. (2017) [9]
SVM 1317 (571+/407-) 5-fold, splitting scheme unknown 0.767 0.948 0.597 0.848 Zhang et al. (2016) [10]
Ensemble of RF and SVM models (5 total) 1241 (683+/558-) 5-fold, splitting scheme unknown 0.701 0.799 0.719 0.909 Ai et al. (2018) [7]
Ensemble of eight different algorithms derived models (8 total) 1254 (636+/618-) 10-fold, splitting scheme unknown 0.783 0.818 0.716 0.773 He et al. (2019) [8]
SVM 401 (174+/227-) 5-fold, Tanimoto similarity based GroupKFold 0.714 ± 0.06 0.697 ± 0.08 0.759 ± 0.03 0.724 ± 0.08 Present study
  1. Literature model performance derived from He et al. (2019) [8]. External test values quoted for the model developed in the present study are for the external test set. Despite being trained on the fewest compounds (401) and using a conservative LOCO-CV cross-validation scheme, the SVM model developed in the present study demonstrated robust predictivity between cross-validation and external test set