Table 2.

Performance comparison of SSM to previous studies on DILIst dataset was retrieved from²⁴

Model	AUC	F1-score	MCC	Accuracy
SSM
SSM - RF (margin to DeepDILI)	0.691 $\pm$ 0.011 (+0.032)	0.784 $\pm$ 0.008 (+0.029)	0.338 $\pm$ 0.030 (+0.007)	0.687 $\pm$ 0.005 (−)
SSM - MLP	0.654 $\pm$ 0.008	0.752 $\pm$ 0.007	0.240 $\pm$ 0.019	0.639 $\pm$ 0.006
SSM - soft voting: RF & MLP	0.664 $\pm$ 0.008	0.760 $\pm$ 0.007	0.264 $\pm$ 0.020	0.683 $\pm$ 0.004
Mold2 descriptor
DeepDILI	0.659	0.755	0.331	0.687
XGBoost	0.651 0.015	0.732 $\pm$ 0.012	0.219 $\pm$ 0.037	0.642 $\pm$ 0.016
RF	0.658 $\pm$ 0.012	0.736 $\pm$ 0.009	0.225 $\pm$ 0.030	0.645 $\pm$ 0.013
SVM	0.645 $\pm$ 0.009	0.752 $\pm$ 0.008	0.220 $\pm$ 0.035	0.646 $\pm$ 0.013
KNN	0.580 $\pm$ 0.021	0.657 $\pm$ 0.020	0.125 $\pm$ 0.038	0.582 $\pm$ 0.019
LR	0.628 $\pm$ 0.009	0.744 $\pm$ 0.007	0.130 $\pm$ 0.038	0.617 $\pm$ 0.011
Deep graph neural network methods
InfoMax	0.624 $\pm$ 0.009	0.687 $\pm$ 0.007	0.226 $\pm$ 0.027	0.627 $\pm$ 0.011
ContextPred	0.628 $\pm$ 0.009	0.687 $\pm$ 0.030	0.242 $\pm$ 0.029	0.632 $\pm$ 0.018
EdgePred	0.642 $\pm$ 0.010	0.690 $\pm$ 0.021	0.261 $\pm$ 0.025	0.639 $\pm$ 0.015
AttrMask	0.608 $\pm$ 0.009	0.653 $\pm$ 0.032	0.203 $\pm$ 0.032	0.606 $\pm$ 0.022
MolHGCN	0.541 $\pm$ 0.024	0.669 $\pm$ 0.023	0.087 $\pm$ 0.051	0.576 $\pm$ 0.025
GraphLOG	0.577 $\pm$ 0.017	0.751	0.000	0.602

Standard error of DeepDILI was not provided from the original article.

Performance without errors in GraphLOG indicates that all predicted values were DILI-positive.

The performance values of the previous models on DILIst data were built on Mold2 descriptor. Performance comparison on TDC-benchmark DILI dataset³⁶ is provided in Table S1.