Table 2.

Dependence of ROC-AUC of RF classifier for general PPIM prediction on various types of unbiased splitting of the data into training and test set (A) Training:Test = 1:1 (B) Training:Test = 3:1

A
Tanimoto cutoff for clustering	AUC-CV (random splitting)	AUC-blind testing (random split)	AVE bias (random split)	AUC-CV (AVE split)	AUC blind testing (AVE split)	AUC-CV realistic-split	AUC blind testing realistic-split
0.90	0.91	0.90	0.28	0.94	0.83	0.96	0.71
0.80	0.85	0.86	0.19	0.91	0.71	0.91	0.78
0.70	0.78	0.80	0.12	0.88	0.66	0.82	0.77
0.60	0.74	0.76	0.07	0.81	0.68	0.78	0.75
B
Tanimoto cutoff for clustering	AUC-CV (random splitting)	AUC-blind testing (random split)	AVE bias (random split)	AUC-CV (AVE split)	AUC blind testing (AVE split)	AUC-CV realistic-split	AUC blind testing realistic-split
0.90	0.92	0.92	0.32	0.94	0.71	0.97	0.56
0.80	0.88	0.87	0.22	0.93	0.68	0.92	0.67
0.70	0.82	0.84	0.13	0.86	0.69	0.86	0.73
0.60	0.77	0.77	0.09	0.82	0.71	0.84	0.64