Application of machine learning methods to histone methylation ChIP-Seq data reveals H4R3me2 globally represses gene expression

Table 6 ML model knockout analysis

Multilinear model knockouts	log2 fold change (predicted WT/predicted KO)
H4R3me2	-0.782
H3R2me1	-0.394
H3K27me2	-0.235
H3K9me1	-0.183
H3K4me3	0.108
H3K4me2	0.285
H3K79me3	0.344
H4K20me1	0.359
H3K79me1	0.428
H3K36me3	0.546
H3K27me2-H3R2me1	-1.192
H3R2me1-H4R3me2	-1.175
H3K27me2-H4R3me2	-1.017
H3K9me1-H4R3me2	-0.964
H3K36me1-H4R3me2	-0.912
H3K79me2-H4R3me2	-0.859
H3K27me3-H4R3me2	-0.837
H3K4me2-H3K36me3	0.857
H3K79me1-H3K79me3	0.903
H3K36me3-H4K20me1	0.907
H3K36me3-H3K79me3	0.916
H3K36me3-H3K79me1	0.976

The log₂ fold change (predicted WT/predicted KO) in average gene expression for single and double knockouts in the multilinear model. In silico knockouts were performed by setting mark amplitudes to zero while fixing all other marks at their experimental values and making model predictions for each gene. The top 5 most repressive and activating fold changes for single and double knockouts are shown. Rows are sorted according to log₂ fold change for single and double knockouts separately.

ISSN: 1471-2105