dCaP: detecting differential binding events in multiple conditions and proteins

Abstract

Background: Current ChIP-seq studies are interested in comparing multiple epigenetic profiles across several cell types and tissues simultaneously for studying constitutive and differential regulation. Simultaneous analysis of multiple epigenetic features in many samples can gain substantial power and specificity than analyzing individual features and/or samples separately. Yet there are currently few tools can perform joint inference of constitutive and differential regulation in multi-feature-multi-condition contexts with statistical testing. Existing tools either test regulatory variation for one factor in multiple samples at a time, or for multiple factors in one or two samples. Many of them only identify binary rather than quantitative variation, which are sensitive to threshold choices.

Results: We propose a novel and powerful method called dCaP for simultaneously detecting constitutive and differential regulation of multiple epigenetic factors in multiple samples. Using simulation, we demonstrate the superior power of dCaP compared to existing methods. We then apply dCaP to two datasets from human and mouse ENCODE projects to demonstrate its utility. We show in the human dataset that the cell-type specific regulatory loci detected by dCaP are significantly enriched near genes with cell-type specific functions and disease relevance. We further show in the mouse dataset that dCaP captures genomic regions showing significant signal variations for TAL1 occupancy between two mouse erythroid cell lines. The novel TAL1 occupancy loci detected only by dCaP are highly enriched with GATA1 occupancy and differential gene expression, while those detected only by other methods are not.

Conclusions: Here, we developed a novel approach to utilize the cooperative property of proteins to detect differential binding given multivariate ChIP-seq samples to provide better power, aiming for complementing existing approaches and providing new insights in the method development in this field.

Figure 1

Flow chart of dCaP.

Figure 2

The ROC curves generated using simulated data at different levels of binding strengths (strong, medium, weak). (A) ROC curves of dCaP-T1-all and dCaP-T1-single. (B) ROC curves of dCaP-T2-all and MANOVA (C) ROC curves of dCaP-T2-pair, DIME and ANOVA.

Figure 3

Length distribution of cell-type specific enriched binding (+) and depleted binding (-) segments. The symbol size is proportional to the frequency of the corresponding segment length.

Figure 4

Patterns of binding and CNV in the five groups of cell-type specific regions identified by dCaP. (A) The five heatmaps represent the five groups of cell-type specific regions identified by dCaP. Each heatmap corresponds to one specific cell-type. Each row in a heatmap shows a cell-type specific region and each column shows the relative binding signals from one of the 15 data tracks. The first three columns are the relative binding signals of the three factors (CTCF, DHS and POL2) in the GM12878 cell lines. The remaining columns are ordered in the same way for HUVEC, HeLa-S3, HepG2 and K562 cell lines, respectively. The rows are first ordered by their clustering class (enriched/depleted) and within each class, the cell-type specific regions are ordered by their genomic locations. (B) Heatmap of CNVs in the five groups of cell-type specific regions. The rows of the heatmap are arranged in the same way as in A and the columns show different types of CNVs in order of GM12878, HUVEC, HeLa-S3, HepG2 and K562. (C) Scatter plots of K-means clustering on the relative binding signals of the three factors in each group of cell-type specific regions.

Figure 5

Examples of cell-type specific regions detected by dCaP. (A) K562-specific depleted binding detected by dCaP (shown in olive) associated with large homozygous deletions on chromosome 9. (B) HeLa-specific depleted binding detected by dCaP close to the TSS of gene STK11 (binding windows shown in purple, differential binding in green and cell-type specific windows in olive).

Figure 6

Functional enrichments from GREAT in the enriched binding cluster of cell-type specific windows.

Figure 7

Comparison of differential binding detected by intersection of binary peaks and dCaP. (A) Venn diagram of TAL1 DPs in the three categories. (B) Scatter plot of the log2 TAL1 binding signals in the merged binding peaks (C) Boxplot of the absolute binding difference in G1E and G1E-ER4 cell lines for the three groups of DPs.