Exploring the ability of machine learning-based virtual screening models to identify the functional groups responsible for binding

Table 2 Performance of different RF_PLEC models when trained on the Polar_ZINC dataset

Model	Accuracy	AU-PRC	Attribution AUC
RF_Morgan	0.52	0.53	0.47
RF_PLEC_2.5	0.62	0.66	0.57
RF_PLEC_3	0.70	0.78	0.65
RF_PLEC_3.5	0.79	0.89	0.77
RF_PLEC_4	0.95	0.99	0.89
RF_PLEC_4.5	0.81	0.86	0.86
RF_PLEC_5	0.79	0.80	0.78
RF_PLEC_5.5	0.75	0.79	0.73
RF_PLEC_6	0.72	0.77	0.70
PointVS	0.89	0.95	0.85

Accuracy denotes the proportion of correctly classified examples on the Polar_ZINC test set, AU-PRC denotes the area under the Precision-Recall curve on the Polar_ZINC test set, and Attribution AUC reflects the ability of a model to correctly identify the ligand atoms responsible for binding on the PDBBind test set (see "Methods" section)
The best performing model was RF_PLEC_4, which uses the same distance cutoff as the Polar deterministic binding rule

ISSN: 1758-2946