Machine intelligence-driven framework for optimized hit selection in virtual screening

Table 4 Performance comparison of PS-driven DNNs framework with other ML algorithms: the comparison of benchmark performance of the PS-driven DNNs/DL framework for hit/lead optimization employing PS module

Algorithm	Dataset	Accuracy	Sensitivity	Specificity	AUC-ROC
Random Forest (RF)	Internal test set (x')^a	0.802	0.754	0.821	0.801
	Small independent validation dataset^b	0.614	0.724	0.488	0.822
	Large independent benchmark dataset^c	0.726	0.817	0.827	0.834
Extreme Gradient Boost (XGB)	Internal test set (x')^a	0.806	0.786	0.813	0.812
	Small independent validation dataset^b	0.631	0.763	0.534	0.699
	Large independent benchmark dataset^c	0.782	0.838	0.621	0.848
Deep Neural Networks (DNNs/DL)	Internal test set (x')^a	0.818	0.913	0.824	0.812
	Small independent validation dataset^b	0.859	0.872	0.822	0.884
	Large independent benchmark dataset^c	0.899	0.902	0.924	0.898

^aTotal of 175 (81 inhibitors and 94 non-inhibitors) partitioned into 7:3 classified as Internal test set (x')
^b46 (35 inhibitors and 11 non-inhibitors) classified as small independent validation dataset
^c1886 (86 inhibitors and 1800 decoys (termed as non-inhibitors)) classified as large independent benchmark dataset

ISSN: 1758-2946