Workflow of MAnorm. MAnorm takes the coordinate of all peaks and aligned reads in both samples as input. The (M, A) value of each common peak is then calculated and plotted, where M = log₂ (Read density in sample 1/Read density in sample 2) and A = 0.5 × log₂ (Read density in sample 1 × Read density in sample 2). Robust regression is subsequently applied to the (M, A) values of all common peaks and a linear model is derived. Finally, the linear model is extrapolated to all peaks for normalization. A P-value is also calculated for each peak to describe the statistical significance of read intensity difference between the two samples being compared.

Normalization of H3K4me3 ChIP-Seq data in H1 ES cells and K562 cells. (a) Venn diagram representing the overlap of H3K4me3 peaks between H1 ES and K562 cells. The overlap of peaks between the two cell lines was 24-fold greater than that observed for random permutations of the peak sets. (b,c) MA plots of all peaks from both samples before (b) and after MAnorm (c). The red line is the linear model derived from common peaks by robust regression. Purple and green circles represent unique peaks; red and black circles represent common peaks. (d) P-values associated with normalized peaks, displayed as an MA plot, with the color range representing -log₁₀ P-value. Most peaks associated with |M| > 1 have a P-value < 10^-10.

Quantitative differences in H3K4me3 marks between two cell lines are strongly correlated with cell type-specific expression of peak targets. (a) Enrichment of the target genes of all common H3K4me3 peaks in H1 ES cells and K562 cells in cell type-specifically expressed genes as identified by SAM (see Materials and methods). The target genes were grouped by the M values of nearby peaks and the enrichment scores were calculated as the ratio of overlap between target genes grouped by M value and differentially expressed genes compared to expected overlap at random. (b,c) Enrichment of the the target genes of all unique H3K4me3 peaks in H1 ES cells (b) or K562 cells (c) in cell type-specifically expressed genes.

Hierarchical clustering of the M value and motif scores in all H3K27ac peaks of H1 ES cells and K562 cells. (a,b) Hierarchical clustering was applied to the correlation coefficients of M values (= log2 (Read density in H1 ES/Read density in K562)) or -M values (= log2 (Read density in K562/Read density in H1 ES)) of all H3K27ac peaks identified in H1 ES cells (a) or K562 cells (b), with motif scores determined for 130 JASPAR vertebrate core motifs in the peak regions. Only the motifs significantly enriched in the peaks of either cell type are shown here (enrichment score > 1.2 and Bonferroni corrected P-value < 1e-5 by Fisher exact test). The names of the motifs closely clustered with M value or -M value are colored in red.

Comparison of c-Myc ChIP-Seq data between HeLaS3 and K562 cell lines. (a) Venn diagram showing the overlap of c-Myc binding site peaks between HeLaS3 and K562 cell lines. The overlap of cMyc peaks between the two cell lines was 65-fold greater than that observed for random permutations of the peak sets. (b,c) Hierarchical clustering of correlation coefficients of M value or -M value of all c-Myc peaks in HeLaS3 cells (b) and K562 cells (c) with the motif scores in the corresponding peak regions. Only significantly enriched motifs are shown. (d) Scatter plot of the M values determined for c-Myc binding versus the M values for H3K27ac based on ChIP-Seq comparisons between HeLaS3 and K562 cell lines.

Comparison of different normalization models. (a-c) MA plot of H3K27ac peaks in H1 ES cells and K562 cells after normalization by total reads (a), quantile normalization (b) and genome-wide MA plot followed by LOWESS regression (c). The corresponding MA plot based on MAnorm is shown in Supplementary Figure 2a in Additional file 2. (d-g) Scatter plot of log2 expression ratios of target genes between H1 ES cells and K562 cells versus the M values normalized by total reads (d), quantile normalization (e), genome-wide MA plot followed by LOWESS normalization (f), and MAnorm (g). The color bar represents the density of dots in the scatter plot and purple dots represent the outliers separated from the others. (h) Distribution of M values for each normalization method and distribution of log2 expression ratios of non-differentially expressed target genes (fold-change < 1.5). T-statistics and P-values calculated based on one sample Students' t-test comparing to 0 for each normalization method were as follows: MAnorm, t-statistic = -0.55 and P = 0.58 by t-test; total reads normalization, t-statistic = -88 and P < 1E-100; quantile normalization, t-statistic = -140 and P < 1E-100; genome-wide MA, t-statistic = 24 and P < 1E-100. For non-differentially expressed target genes, t-statistic = -0.76 and P = 0.45.