Estimation of the applicability domain of kernel-based machine learning models for virtual screening

Fechner, Nikolas; Jahn, Andreas; Hinselmann, Georg; Zell, Andreas

doi:10.1186/1758-2946-2-2

Journal of Cheminformatics

Table 4 Virtual Screening results for Thrombin at five different applicability levels

From: Estimation of the applicability domain of kernel-based machine learning models for virtual screening

Kernel	ADE		Threshold	AUROC	BEDROC (α)				Ligands	Decoys
					100.00	53.6	32.2	20.0
OAK	KDE	50%	0.64	0.88	0.61	0.60	0.60	0.61	11	575
		33%	0.66	0.87	0.66	0.63	0.63	0.64	10	380
		200	0.68	0.97	0.78	0.72	0.73	0.73	8	192
		100	0.69	0.97	0.90	0.81	0.81	0.78	8	92
	wKDE	50%	0.64	0.88	0.61	0.60	0.60	0.61	11	575
		33%	0.66	0.87	0.66	0.63	0.63	0.64	10	380
		200	0.68	0.97	0.78	0.73	0.73	0.74	8	192
		100	0.69	0.97	0.90	0.81	0.81	0.78	8	92
	ONE	50%	-9.61	0.88	0.62	0.60	0.60	0.60	11	478
		33%	-9.11	0.87	0.66	0.63	0.63	0.64	10	380
		200	-8.06	0.94	0.77	0.69	0.68	0.68	9	191
		100	-7.08	0.97	0.90	0.81	0.81	0.78	8	92
FlexOAK	KDE	50%	0.64	0.85	0.43	0.47	0.48	0.52	11	575
		33%	0.65	0.91	0.49	0.53	0.53	0.59	9	381
		200	0.67	0.91	0.59	0.58	0.58	0.63	8	192
		100	0.68	0.91	0.74	0.67	0.67	0.69	8	92
	wKDE	50%	0.64	0.78	0.41	0.45	0.46	0.49	12	574
		33%	0.65	0.91	0.49	0.53	0.53	0.59	9	381
		200	0.67	0.91	0.59	0.58	0.58	0.63	8	192
		100	0.68	0.91	0.74	0.67	0.67	0.69	8	92
	ONE	50%	-9.80	0.91	0.49	0.54	0.54	0.59	10	576
		33%	-9.07	0.91	0.55	0.59	0.60	0.64	9	381
		200	-8.10	0.92	0.62	0.63	0.64	0.68	8	192
		100	-7.32	0.98	0.74	0.68	0.68	0.71	7	93
MARG	KDE	50%	0.876	0.81	0.45	0.43	0.43	0.45	12	574
		33%	0.885	0.96	0.56	0.58	0.58	0.62	8	382
		200	0.893	0.94	0.62	0.58	0.58	0.59	8	192
		100	0.899	0.93	0.90	0.80	0.79	0.74	8	92
	wKDE	50%	0.882	0.81	0.46	0.44	0.44	0.47	12	574
		33%	0.891	0.96	0.56	0.58	0.58	0.62	8	382
		200	0.899	0.95	0.77	0.69	0.69	0.68	8	192
		100	0.905	0.93	0.90	0.80	0.79	0.74	8	92
	ONE	50%	-2.631	0.80	0.42	0.42	0.43	0.46	11	531
		33%	-2.334	0.88	0.52	0.50	0.50	0.52	9	381
		200	-1.932	0.93	0.61	0.54	0.54	0.54	8	192
		100	-1.607	0.92	0.73	0.64	0.63	0.61	7	93

The threshold is adjusted such that either a certain fraction (50%, 33%) of the compounds or that exactly n compounds (n = 200, 100) remain in the domain. Results, which differ significantly from random rankings (p-Value < 0.01) are shown in bold.

Back to article page

ISSN: 1758-2946

Contact us

Submission enquiries: journalsubmissions@springernature.com