LightCpG: a multi-view CpG sites detection on single-cell whole genome sequence data

BMC Genomics

Table 4 The comparison of different methods for the best evaluation values in all cells

Dataset	Methods	Acc(%)	AUC	Fscore(%)	MCC(%)	SP(%)	SE(%)
HCCs	LightCpG	94.07	0.9709	85.82	82.10	95.74	87.59
	DeepCpG	93.20	0.9732	85.94	81.67	93.94	90.71
	RF Zhang	90.29	0.9388	80.10	73.95	91.78	85.29
HepG2	LightCpG	83.51	0.9246	82.40	68.04	90.33	78.43
	DeepCpG	84.08	0.9239	82.82	68.14	82.85	85.60
	RF Zhang	81.25	0.8954	80.05	63.48	87.77	76.35

¹LightCpG employs three types of features (sequence feature, structural feature and positional feature) and LightGBM [36] to identify the CpG sites
²DeepCpG [35] embodies the connection between various cells by using the deep learning model Gated Recurrent Network (GRU) and also extracts features from the DNA sequence by the convolutional neural network (CNN) and one additional fully connected hidden layer. Then DeepCpG uses Fully Connected Neural Network to identify CpG sites
³RF Zhang [1] extracts the genomic positional features, neighbor features, sequence properties and sic-regulatory elements to identify the CpG sites
⁴The boldface is the best value in the column

ISSN: 1471-2164