Table 1.

Performance results of PredPS and existing molecular representation methods on the internal dataset.

	RF	SVM	DMPNN	CMPNN	PredPS w/o FP	PredPS w/ FP
Accuracy	0.766 ± 0.005	0.734 ± 0.010	0.795 ± 0.013	0.815 ± 0.025	0.807 ± 0.022	0.835 ± 0.007
Sensitivity	0.741 ± 0.039	0.745 ± 0.026	0.767 ± 0.073	0.769 ± 0.027	0.802 ± 0.059	0.823 ± 0.054
Specificity	0.791 ± 0.031	0.726 ± 0.038	0.823 ± 0.055	0.855 ± 0.052	0.813 ± 0.046	0.846 ± 0.049
AUC	0.817 ± 0.008	0.804 ± 0.011	0.873 ± 0.009	0.881 ± 0.002	0.897 ± 0.002	0.901 ± 0.006

Random forest (RF) and support vector machine (SVM) were implemented using Scikit-learn package. The DMPNN was implemented with the source code obtained from ChemProp (https://github.com/chemprop/chemprop), and the CMPNN was implemented with the source code obtained from https://github.com/SY575/CMPNN. We employed a 5-fold cross-validation with a random split and provided the mean and standard deviation for each performance metric.