Table 2.

Model building strategies and performance of enhancer prediction methods

Category	Method	Operational model	Positive predictive value (%)	Note	Ref
Discriminative model	Heintzman’s method	Thresholds of histone modification profiles	39.5	Mapped to distal p300 binding sites in HeLa cells	[5]
	Visel’s method (2009)	Thresholds of p300 binding profiles	87.7	With reproducible enhancer activity in transgenic mouse	[46]
	Narlikar’s method	Linear regression	62	With reproducible enhancer activity in vivo in mouse and zebrafish	[65]
	Zinzen’s method	Support vector machine	71.4	With reproducible enhancer activity in transgenic Drosophila	[67]
	Firpi’s method	Time-delay neural network	66.3	Overlapped with p300 binding sites, Dnase I hypersensitivity sites or TRAP220 binding sites in HeLa cells	[10]
	Lee’s method	Support vector machine	74.5	Overlapped with Dnase I hypersensitive enhancers in embryonic mouse whole brain cells	[44]
	ChromaGenSVM	Support vector machine	57	Overlapped with p300 binding sites, Dnase I hypersensitivity sites or TRAP220 binding sites in HeLa cells	[9]
Probabilistic graphical model	Won’s method	Hidden Markov model	54.8	Overlapped with p300 binding sites, Dnase I hypersensitivity sites or TRAP220 binding sites in HeLa cells	[11]
	Bonn’s method	Bayesian network	78	Overlapped with previously identified TF binding sites in Drosophila	[6]
Other	Chen’s method	Multinomial logistic	83	Overlapped with at least one TF peak from 7 mouse embryonic stem cell ChIP-seq datasets	[8]
	Yip’s method	Random forest	67	With enhancer activity in vivo in mouse and medaka fish (28/42)	[50]

Note: The performance shown here is the reported performance compared to experimental results. The positive predictive value (percentage) was calculated as follows: positive predictive value = true positive/(true positive + false positive).