Predicting environmentally responsive transgenerational differential DNA methylated regions (epimutations) in the genome using a hybrid deep-machine learning approach

Table 2 DMR prediction performance of hybrid model learned from all training data

Chr	#Predicted DMR	%Recall	#max poss. DMRs	%max poss. DMRs	%Genome
1	127,816	96.09	267,040	95.03	45.54
2	141,619	85.88	250,909	94.68	53.44
3	105,790	92.90	168,257	94.52	60.10
4	159,974	97.34	152,775	95.49	87.41
5	74,363	94.15	164,461	95.61	43.23
6	103,072	93.77	139,443	95.50	70.59
7	76,805	83.48	137,363	95.39	53.33
8	57,713	97.74	127,323	96.45	43.72
9	73,307	96.03	115,863	95.75	60.58
10	93,641	87.19	108,271	97.54	84.36
11	39,174	87.54	85,854	96.46	44.01
12	43,108	91.82	110,497	97.78	84.52
13	44,317	96.31	108,528	95.00	39.21
14	62,311	94.63	104,104	94.64	54.65
15	96,840	74.41	84,065	94.45	88.03
16	62,548	94.86	84,713	95.18	70.27
17	71,651	98.56	83,393	95.85	80.50
18	53,671	98.38	83,408	95.87	61.69
19	51,335	88.20	58,891	96.54	84.18
20	18,205	98.33	47,449	96.83	37.12
X	47,092	68.45	144,335	91.35	2.98
Y	2608	91.13	3159	95.43	85.94
ALL	1,748,888	95.49	2,742,978	95.40	63.75

The number of DMRs in a chromosome predicted by the hybrid model trained on data from that chromosome, and the number of DMRs predicted across the whole genome (ALL) by the hybrid model trained on data from the whole genome (#Predicted DMR). Also shown is the percentage recall (%Recall), which is the percentage of the training DMRs that the model correctly predicts as DMRs. As a comparison, “maximum possible DMRs” is defined as the set of all 1000 bp regions minus those regions that are clearly nonDMRs, because they have no CpGs or more than 20% (200) CpGs. The size of this “maximum possible DMRs” set serves as an upperbound on the number of possible DMRs, and the number of predicted DMRs should be well below this bound. The table shows the size of this set (# max poss. DMRs) and the percentage of the chromosome or whole genome this set represents (% max poss. DMRs). The %Genome column shows the percent of the chromosome, or entire genome for ALL, that the predicted DMRs represent. The %Genome value should be well below the “% max poss. DMRs” value

ISSN: 1471-2105