Table 4.

3F model selection on genotype-phenotype data up to September 2006

		3F Model generation a			3F Model selection b
drug		# 3F Models	# lower SBC	# lower AIC	lower SBC c	lower AIC d	N test	ase	3F model selected e
Nucleoside RT inhibitorsf	AZT	300	86	0	yes	no	800	0.103	296
	3TC	150	60	34	yes	no	807	0.037	99
	ddI	150	20	70	no	yes	807	0.049	83
	d4T	120	41	35	yes	no	806	0.040	81
	ABC	200	111	53	yes	no	807	0.038	95
	FTC	80	28	22	yes	yes	804	0.071	76
	TDF	400	66	196	no	yes	807	0.039	298
NNRTIg	NVP	400	93	0	yes	no	801	0.089	391
	EFV	500	101	0	yes	no	807	0.246	386
	ETR	700	49	0	yes	no	777	0.113	656
Protease inhibitors	IDV	485	50	51	yes	yes	805	0.075	482
	NFV	375	64	6	yes	yes	808	0.063	375
	SQV	600	53	0	yes	no	807	0.092	575
	APV	1000	0	656	no	yes	808	0.060	709
	LPV	500	205	28	yes	no	807	0.157	319
	ATV	1275	0	2	no	yes	805	0.117	1158h
	TPV	1000	641	142	yes	no	806	0.059	428
	DRV	1000	823	799	yes	yes	816	0.096	707

^aThe number of 3F models generated was arbitrary but taken large enough such that at least one 3F model was found with a lower SBC or AIC than the reference on the genotype-phenotype data set up to July 2006.

^bFrom the remaining 3F models with lower SBC or AIC than the reference, the 3F model was then selected with the lowest average squared error (ase) on an unseen genotype-phenotype data set collected between July and September 2006 (test set) containing approximately 800 samples.

^cSBC of the selected 3F model < SBC reference on the test set (yes/no).

^dAIC of the selected 3F model < AIC reference on the test set (yes/no).

^eThe number of different random divisions used in the stepwise regression in the selected 3F model.

^fFor the nucleoside RT inhibitors the number of random divisions needed was less than 100, with exception of AZT and TDF.

^gFor the non-nucleoside RT inhibitors most random divisions were needed for ETR.

^hATV was the only drug for which more than 1000 different random divisions were needed.