Table 1.

Performance of the ensemble models (including the non-machine-learning score-only classifier) under different activity thresholds and machine learning approaches.

Threshold	1 µM	10 µM	100 µM
Number of active ligands
of the 230 in the training set	76 (33%)	107 (46%)	127 (55%)
of the 64 in the test set	19 (30%)	33 (52%)	36 (56%)
Vina score (no statistical model)
Ensemble AUC (training set)	0.54	0.57	0.63
Best AUC among all individual structures	0.58	0.62	0.66
Ensemble MLP model
AUC (std. dev.) 5CV (training set)	0.84 (0.05)	0.81 (0.03)	0.75 (0.06)
AUC (test set)	0.82	0.90	0.92
Matthews coeff. (test set)	0.49	0.56	0.59
Ensemble LDA model
AUC (std. dev.) 5CV (training set)	0.78 (0.10)	0.77 (0.08)	0.81 (0.07)
AUC (test set)	0.77	0.77	0.77
Matthews coeff. (test set)	0.38	0.34	0.43
Ensemble QDA model
AUC (std. dev.) 5CV (training set)	0.79 (0.12)	0.77 (0.08)	0.81 (0.10)
AUC (test set)	0.76	0.78	0.75
Matthews coeff. (test set)	0.31	0.47	0.39