Table 3. Performances of Models Developed Using Different Descriptor Selection Procedures^a.

	unsupervised selection			neural network pruning
		RMSE			RMSE
descriptor set	N	training	test	N	training	test
CDK	159	0.93	1.13	6	0.89	1.2
Dragon	1824	0.93	1.15	18	0.87	1.19
Fragmentor	631	0.98	1.18	12	0.92	1.21
GSFrag	202	0.97	1.1	24	0.97	1.18
Mera, Mersy	242	0.93	1.04	10	0.93	1.18
Chemaxon	97	0.93	1.16	11	0.92	1.16
Inductive	39	0.94	1.17	21	0.93	1.16
Adriana	133	0.93	1.14	8	0.92	1.1
QNPR	381	0.95	1.12	74	0.89	1.13
E-state	185	0.96	1.16	11	0.9	1.24
Consensus	4036	0.88	1.08	186	0.85	1.13

^aN is the number of descriptors selected to develop the respective model. RMSE is the root mean squared error calculated for the training (n = 483) and full test set (n = 143).