Data normalization is an important step in sequencing data analysis. However, normalization of ChIP-Seq data is a difficult task due to the differential S/N ratio across samples (see Discussion). These differences cannot simply be addressed using traditional microarray data normalization methods, such as quantile normalization 14 and MA plot followed by LOWESS regression 13. Here we borrow the idea of the MA plot and propose a novel method for quantitative comparison of ChIP-Seq data sets based on two empirical assumptions. First, we assume the true intensities of most common peaks are the same between two ChIP-Seq samples. This assumption is valid when the binding regions represented by the common peaks show a much higher level of co-localization between samples than that expected at random, and thus binding at the common peaks should be determined by similar mechanisms and exhibit similar global binding intensity between samples. Second, the observed differences in sequence read density in common peaks are presumed to reflect the scaling relationship of ChIP-Seq signals between two samples, which can thus be applied to all peaks. Based on these hypotheses, the log₂ ratio of read density between two samples (M) was plotted against the average log₂ read density (A) for all peaks, and robust linear regression was applied to fit the global dependence between the M-A values of common peaks. Finally, the derived linear model was used as a reference for normalization and extrapolated to all peaks. The normalized M value was then used as a quantitative measure of differential binding in each peak region between two samples, with peak regions associated with larger absolute M values exhibiting greater differences in binding. The workflow of the method, MAnorm, is shown in Figure 1. The MAnorm package is available for download in Additional file 1.

Differential epigenetic modifications are closely associated with many developmental and disease processes 15. As such, quantitative comparison of ChIP-Seq signals across multiple cell types may help elucidate underlying epigenetic mechanisms of disease and tissue-specific regulation. We applied MAnorm to analyze the differences between H1 human embryonic stem (ES) cells and two disease-related cell lines, K562 and HeLaS3, for two histone modifications positively associated with gene expression, H3K4me3 and H3K27ac. For each chromatin mark, peaks identified in each cell line showed substantial overlap with those from the other two cell lines, with the overlap ranging from 16- to 24-fold greater than the overlap observed by random permutations (Figure 2a; Supplementary Figure 1 in Additional file 2). Before normalization, the MA plots exhibited an overall global dependence of M value on A, which was closely fitted by a linear model derived by robust regression (Figure 2b; Supplementary Figure 2 in Additional file 2). A similar global dependence was evident in comparisons of biological replicates (Supplementary Figure 7 in Additional file 2; discussed below), indicating the dependence of M on A does not reflect biological changes but is due mainly to systemic bias and noise. After application of MAnorm to remove this dependence from the set of common peaks, the distribution of common peaks became highly symmetric with respect to the new A axis. Furthermore, the two sets of unique peaks became more symmetric in all comparisons (Figure 2c; Supplementary Figure 2 in Additional file 2). These observations suggest that the ChIP-Seq signals in all peaks follow a similar scaling relationship and that the extrapolation of the linear model from common peaks to all peaks is valid. The significance of differential binding in each peak region was determined using a P-value calculated based on a Bayesian model developed by Audic and Claverie 16 (Figure 2d; Supplementary Figure 2 in Additional file 2).

Next, we investigated the relationship between the M value (= log₂ (Read density in cell type 1/Read density in cell type 2)) and the change in expression of peak targets between cell types. In general, target genes associated with positive M values - that is, peaks with higher H3K4me3 and H3K27ac read intensity in cell type 1 - were enriched in genes more highly expressed in cell type 1. Conversely, target genes associated with negative M values were enriched in genes more highly expressed in cell type 2 (Figure 3; Supplementary Figure 3 in Additional file 2). These findings are consistent with the activating role of these two histone modifications 17. Notably, the enrichment score of genes more highly expressed in cell type 1 showed strong positive correlation with the M values, while the enrichment score of genes more highly expressed in cell type 2 correlated negatively with M, suggesting that the M statistics determined by MAnorm serve as an indicator of cell type-specificity for the epigenetic marks in peak regions (Figure 3; Supplementary Figure 3 in Additional file 2). Furthermore, the target genes associated with an absolute M value > 1 were significantly enriched in genes highly expressed in the corresponding cell type among all our comparisons, implying that the absolute M value of 1 is a suitable cutoff for defining cell type-specifically marked genes. It should be noted that many common target genes were associated with M values far from 0, and were still highly enriched for cell type-specifically expressed genes (Figure 3a; Supplementary Figure 3a in Additional file 2), indicating that the differential epigenetic marks at these genes are also functional. On the other hand, those unique target genes with M values near zero displayed much weaker enrichment of cell type-specifically expressed genes (Figure 3b, c; Supplementary Figure 3b, c in Additional file 2), indicating that they are not uniquely marked in one cell type. MAnorm also exhibited good performance when applied to ChIP-seq datasets composed of broad, diffuse peaks, such as histone modifications like H3K36me3 (Supplementary Figure 4 in Additional file 2 and Supplementary Text in Additional file 3). In conclusion, MAnorm quantitatively describes authentic binding differences of chromatin-associated proteins, and thus represents an improvement over arbitrary definitions of common and unique targets based on peak overlap between samples.

A conventional strategy to identify cell type-specific regulators associated with changes in epigenetic marks relies on the identification of transcription factor binding sites that are highly enriched in unique peak regions. This method often yields multiple candidates, and thus complicates the identification of key regulators associated with the differences in epigenetic marks in each cell type. One advantage of the continuous M value determined by MAnorm is that it can be used to identify potential regulators driving cell type-specific epigenetic modifications. To do so, we searched for motifs that show the highest correlation with M values for all peaks. For example, we compared H1 ES and K562 cell lines for differences in H3K27ac, a histone mark that serves as an indicator of both active promoters and cell type-specific enhancers 1819. We found that OCT4 (POU5F1) and SOX2 binding motifs were closely clustered with the M value (= log₂ (H3K27ac read density in H1 ES cells/H3K27ac read density in K562 cells) of H3K27ac peaks (Figure 4a), suggesting the corresponding factors are closely related to the activation of ES cell-specific genes and cis-elements. In contrast, -M value (= log₂ (H3K27ac read density in K562 cells/H3K27ac read density in H1 ES cells) formed a compact module with the binding motifs for transcription factors GATA1 and SCL (TAL1) (Figure 4b), suggesting their roles as regulators favoring H3K27ac modification in K562 cells. These findings are consistent with the established roles of OCT4-SOX2 in ES cell self-renewal 2021 and GATA-SCL in hematopoiesis and leukemogenesis 5. On the other hand, several motifs, including MYC and ETS motifs (for example, ELK1, ELK4, GABPA), were highly enriched in both peak sets, but showed no association with the differential binding of H3K27ac (specifically, M value); this indicates they are involved in H3K27ac modification in a non-cell type-specific manner. This finding in turn supports the working assumption of our model that binding at most common peaks is determined by similar mechanisms. Furthermore, upon comparison of the H3K27ac marks in H1 ES or K562 cells with those in HeLaS3 cells, these same transcription factor motifs were tightly associated with the H1 ES or K562-specific enrichment of H3K27ac marks at the corresponding target regions (Supplementary Figure 5 in Additional file 2), indicating the clustering results are robust. Thus, MAnorm serves as a powerful tool to uncover transcription factor motifs and factors critical for cell-specific gene regulation.

The oncogene Myc (c-Myc) is an important transcriptional regulator in both ES cells and cancer cells 2223. Mechanisms underlying its cell type-specific binding are largely unknown. We applied MAnorm to quantify differential binding of c-Myc in HeLaS3 and K562 cells and explored its relationship with other factors. Using a P-value cutoff of 1E-6, 18,924 peaks were detected in the c-Myc ChIP-Seq data set of HeLaS3 cells, and 13,140 peaks were detected in K562 cells; approximately 6,000 peaks were common to both cell lines (Figure 5a). MAnorm largely removed the global dependence of M on A (Supplementary Figure 6a, b in Additional file 2). A significant fraction of c-Myc peaks were associated with M values far from zero, suggesting that c-Myc has a large number of differential binding loci between HeLaS3 cells and K562 cells. To search for cell line-specific co-factors that might contribute to such differential binding, we performed hierarchical clustering between the M statistics inferred by MAnorm and the motif scores in the c-Myc binding peaks. The c-Myc motif was highly enriched in both sets of c-Myc peaks (data not shown), but did not show significant correlation with M statistics in either clustering map (Figure 5b, c), indicating that the c-Myc motif is not responsible for the cell line-differential binding seen for c-Myc. Of note, the M statistic (= log₂ (c-Myc read density in K562/c-Myc read density in HeLaS3)) clustered with the motifs of two other factors, GATA1 and SCL (TAL1) (Figure 5b), and the -M statistic (= log₂ (c-Myc read density in HeLaS3/c-Myc read density in K562) clustered with the motifs of AP1 and TEAD1 (Figure 5c). Strikingly, these clustering patterns were highly similar to those obtained from the comparison of the H3K27ac mark between these two cell types (Supplementary Figure 5b in Additional file 2), suggesting an underlying correlation between the cell type-specific binding of c-Myc and the H3K27ac mark. To test whether this was the case, we mapped c-Myc binding sites to gene promoters, and found that for the 9,013 genes targeted by both c-Myc and H3K27ac, the Pearson correlation coefficient between the M statistics of c-Myc and H3K27ac was 0.73 (Figure 5d), lending further support to our clustering result.

Integrating ChIP-Seq data from multiple biological replicates, which in some cases are generated by different laboratories and/or using different platforms, may be employed to reduce the false positive rate in identified binding sites. A simple approach is to define a stringent set of peaks composed only of the common peaks shared by two or more replicates. However, this method is highly sensitive to peak cutoff and may exclude peaks that have similar ChIP intensities between replicates. Moreover, some common peaks that show dramatic differences in read density are retained. Therefore, to make full use of the information in biological replicates, a quantitative comparison of peak intensity is particularly useful. We have applied MAnorm to compare two replicates of H1 ES cell H3K27ac ChIP-Seq data. After application of MAnorm (Supplementary Figure 7a, b in Additional file 2), many of the unique peaks were associated with M values close to zero, indicating that these peaks exhibit good reproducibility between replicates. On the other hand, there remained a small fraction of common peaks with M values far from zero, representing strong signal differences between replicates. Next, we showed that the M value between replicates is a good indicator of H3K27ac target gene expression. We grouped H3K27ac target genes by the absolute value of M statistics and calculated the expression distribution of each gene group. Given that H3K27ac marks are positively associated with gene expression, we anticipated that more highly expressed genes will have stronger H3K27ac marks, and therefore be more reliable. In fact, we observed that genes having higher expression tend to be the targets of H3K27ac peaks with lower absolute M values, that is, peaks showing smaller difference between replicates, for both common peaks and unique peaks (Supplementary Figure 7c-e in Additional file 2). Furthermore, by overlapping the above set of ENCODE peaks with H3K27ac peaks for H1 ES cells generated in a different laboratory 19, we found that a much lower proportion of the peaks with |M| > 1 were covered by the new peak set than those with |M| < 1 (Supplementary Figure 7f in Additional file 2). This suggests that |M| = 1 can also be used as an empirical cutoff to filter unreliable peaks. Thus, MAnorm can be used both to check whether two replicates are concordant, and also to obtain high confidence peak lists by filtering out inconsistent peaks. Compared with arbitrary removal of unique peaks, MAnorm allows for better use of replicate peak data. The MAnorm package (Additional file 1) provides the opportunity to list concordant and non-concordant peaks between two samples based on user-specified cutoffs, with the concordant peak list corresponding to high-confidence peaks.

We compared the performance of MAnorm with three widely used normalization methods that use genome-wide signals as reference, namely, normalization by total reads, quantile normalization, which assumes the genome-wide distribution of read densities to be the same across samples, and normalization using a genome-wide MA plot followed by LOWESS regression. We used all four methods to compare H3K27ac ChIP-Seq data between H1 ES and K562 cells. The MA plot normalized by MAnorm (Supplementary Figure 2a in Additional file 2) was relatively symmetric, while corresponding plots obtained by the other three normalization methods remained highly asymmetric. Of note, the common peaks showed a clear global bias towards stronger binding in K562 cells for total read normalization and quantile normalization (Figure 6a, b) and toward H1 ES cells for genome-wide MA plot normalization (Figure 6c). To examine which method better reflects a true biological signal, we compared M values normalized by all four methods with the expression change of target genes. If a specific type of histone modification is closely related to gene regulation, the direction of histone modification change should be consistent with that of the change in expression of the target genes. By visual inspection, we found this was true for the M values normalized by MAnorm (Figure 6g). In contrast, M values normalized by the other three methods were inconsistent with the log2-expression ratios of target genes (Figure 6d-f). Specifically, most of the genes with no change in H3K27ac levels (M = 0) had higher (total read and quantile normalization) or lower (genome-wide MA plot normalization) expression in H1 ES cells compared to K562 cells; while the majority of the genes expressed at similar levels in these two cell types were associated with negative (total read and quantile normalization) or positive (genome-wide MA plot normalization) M values, that is, they had higher (total read and quantile normalization) or lower (genome-wide MA plot normalization) levels of H3K27ac in K562 cells.

To quantitatively measure the bias of the M values given by the above normalization methods, we first collected non-differentially expressed genes (fold-change < 1.5) between H1 ES cells and K562 cells. As shown in Figure 6h, these genes are indeed not differentially expressed (t-statistics = -0.76 and P-value = 0.45 by Students' t-test in comparison to an expression ratio of 1 (M = 0)), indicating they are suitable for our comparison. Since H3K27ac marks are closely associated with transcriptional activation, it is reasonable to assume that these non-differentially expressed genes should exhibit similar global H3K27ac levels. This is true only for H3K27ac levels determined by MAnorm, where the M values for H3K27ac of the non-differentially expressed target genes were not significantly different from a ratio of 1 (M = 0; t-statistic = -0.55 and P-value = 0.58 by t-test; Figure 6h, red curve). In contrast, M values for H3K27ac obtained by the other three normalization methods exhibited large deviations from M = 0 (t-statistic ranging from 24 to 140 and P-value < 1e-100; Figure 6h). Thus, MAnorm exhibits superior performance in identifying authentic biological changes.

We also compared the performance of MAnorm in detecting differential binding regions in ChIP-Seq data sets with that of two currently used statistical methods, ChIPdiff 11 and MACS 4. For this analysis, one data set was used as sample and the other was used as control in order to detect regions with significantly elevated ChIP-Seq signals in the first data set 4. We applied all three methods to compare ChIP-Seq data for H3K27ac marks between H1 ES cells and K562 cells (Supplementary Table 1 in Additional file 4). ChIPdiff and MACS identified four to six times more target regions associated with significantly increased ChIP-Seq signals for K562 cells compared with those found for H1 ES cells, whereas MAnorm yielded a similar number of cell type-biased peaks in each cell line. To compare the enrichment of cell type-specifically expressed genes in the sets of target genes of the differential binding regions discovered by the three methods, we selected the same number of target genes associated with top differential binding regions identified by each method. The target genes of top differential binding regions identified by MAnorm contained similar numbers of H1 ES cell highly expressed genes but a greater number of K562 cell highly expressed genes compared to those identified by ChIPdiff and MACS (Supplementary Table 1 in Additional file 4), suggesting MAnorm performs better in detecting differentially binding regions than the other two methods. Importantly, the fold changes of differential binding given by ChIPdiff and MACS were based on the total number of reads, which may not be appropriate, as discussed above. Additionally, MAnorm showed even better enrichment of cell type-specifically expressed genes in differential binding region targets than the method developed by Taslim et al. 12 when applied to ChIP-Seq data presented in their study (Supplementary Table 2 in Additional file 4).

The performance of MAnorm was tested using ENCODE ChIP-Seq data describing histone modifications (H3K4me3 and H3K27ac) 34 and transcription factor binding (c-Myc and Pol II) 35 across three human cell lines: H1 ES cells, HeLaS3 cells, and K562 cells 36. Since these data were generated and processed by different laboratories associated with the ENCODE project, the data sets were reanalyzed and the ChIP-Seq peaks in each sample were redefined using MACS 4 using a P-value cutoff of 1e-10 for histone modifications and a P-value cutoff of 1e-6 for transcription factor binding. The peaks of histone modifications were further filtered by the false discovery rate (FDR) values modeled by MACS. The target genes of each group of peaks were defined as those RefSeq genes that have a given peak(s) in the promoter region, defined as the region from 8 kb upstream to 2 kb downstream of the transcription start site.

Gene expression data for all three cell types were collected from the Gene Expression Omnibus (GEO) database using accession numbers [GEO:GSE26312] (for H1 ES cells) 29, [GEO:GSE2735] (for HeLaS3 cells) 37 and [GEO:GSE12056] (for K562 cells) 38, and the raw data were reprocessed by dChip 39. The differentially expressed genes were subsequently identified by Significance Analysis of Microarrays (SAM) 40 using a combined cutoff of fold change > 2 and FDR < 0.01. In total, 3,465 genes more highly expressed in H1 ES cells and 2,224 genes more highly expressed in K562 cells were identified from the H1 ES to K562 comparison; 5,815 genes more highly expressed in H1 ES cells and 1,649 genes more highly expressed in HeLaS3 cells were identified from the H1 ES cell to HeLaS3 cell comparison; and 3,555 genes more highly expressed in HeLaS3 cells and 5,916 genes more highly expressed in K562 cells were identified from the HeLaS3 cell to K562 cell comparison. To study the relationship between binding differences in peak regions and the expression change of the corresponding target genes, we used the M values of peaks to divide the targeted genes into different groups separated by integer M values from -4 to 4, and then calculated the enrichment score of the overlap between each gene group and those differentially expressed genes. To avoid extreme enrichment scores, groups composed of < 50 genes were merged with the larger of the adjacent two gene groups.

To detect the potential binding of transcription factors in defined peak regions, we downloaded the position weight matrixes of 130 core vertebrate motifs from the JASPAR database 41 and performed motif scan 42 applied to a 1,000 bp window centered at the peak summit. For each motif F, the raw motif matching score at each peak P was calculated as:

max s ∈ p log P ( S | M ) P ( S | B )

in which S is a sequence fragment of the same length as the motif and B is the background frequency of different nucleotides estimated from 10,000 random 1,000 bp sequences sampled from the genome. The motif score of motif M in peak P was defined as the raw motif matching score divided by the maximum possible score, that is, the raw motif score obtained by the consensus sequence of the motif.

To identify transcription factors associated with cell type-specific binding of the ChIP'd proteins, we applied hierarchical clustering with Ward's linkage to cluster the M value with the motif matching score of JASPAR motifs in all peaks of cell type 1, and separately the -M value was clustered with the motif scores in all peaks of cell type 2, using '1 - ρ' as the distance metric, where ρ is the Pearson correlation coefficient. Only motifs with an enrichment score > 1.2 and Bonferroni-corrected P-value < 1.0E-5 by Fisher exact test are shown in the clustering plots.

For total read normalization, we divided the read intensity of each peak region by the total number of mapped sequence reads. For quantile normalization, we first divided the whole genome into non-overlapping bins of the same size as the window used in MAnorm (2,000 bp for H3K27ac), and then calculated the read count in each bin. Finally, the distribution of bin read counts was normalized to be the same by matching all quantiles between samples. For normalization by genome-wide MA plots, we first divided the whole genome into non-overlapping bins of the same size as the window used in MAnorm (2,000 bp for H3K27ac), and then calculated the M-A value of each bin. The dependence between M-A value was then removed by subtracting M values with local linear model fitted by LOWESS regression from the genome-wide M-A values.

To compare the performance of MAnorm with the model developed by Taslim et al. 12, we used MACS to identify peaks from the same Pol II ChIP-Seq datasets used by 12, and then applied MAnorm to compare Pol II binding profiles between estradiol (E2)-stimulated MCF7 cells and unstimulated MCF7 cells. The gene expression data of unstimulated and E2-stimulated MCF7 cells was obtained from the GEO database, accession number [GEO:GSE11352] 43. We identified 59 genes showing higher expression in unstimulated MCF7 cells and 130 genes showing higher expression in E2-stimulated (12 h) MCF7 cells using SAM with fold change > 2 and FDR < 0.1. Finally, the performance of MAnorm was evaluated by comparing the difference of Pol II binding determined by both models with the differential expression of target genes.