Chip-seq data analysis: from quality check to motif discovery and more

Lausanne, 4 - 8 April 2016

ChIP-partitioning tool: shape based analysis of transcription factor binding tags

Romain Groux, Sunil Kumar and Philipp Bucher

PDF and MS Word version of the tutorials.


Probabilistic partitioning methods to discover significant patterns in ChIP-Seq data will be discussed in this section with multiple simulated and biological examples [Nair et al., 2014]. Our methods take into account signal magnitude, shape, strand orientation and shifts. We have compared this methods with some of the existing methods and demonstrated significant improvements, especially with sparse data [please refer the publication for details on comparative analysis]. Besides pattern discovery and classification, probabilistic partitioning can serve other purposes in ChIP-Seq data analysis. In this section we will also will exemplify its merits in the context of peak finding and partitioning of ChIP-seq tag distribution for transcription factor binding (ex: CTCF).

Hints and recipes

The exercise is divided into four parts. First part is using simulated data and other three parts on real experimental data. Before creating the input data, lets define ChIP partitioning functions.

Definition of function

Various variants of Expectation-Maximization partitioning methods can be downloaded form here: EM functions. Download and save the file as em_functions.R.

ChIP partitioning using different datasets

Part I: Simulated data [To be performed in R]

Part II: CTCF dataset [GEO series GSE29611 Histone Modifications by ChIP-seq from ENCODE/Broad Institute]

Part III: Nucleosome organization around transcription factor binding sites in GM12878 cells. For this exercise, the datasets come from the following sources : ENCODE Uniform, GSE32970, GSE35586, GSE29611


Here are the needed R functions em_api.R. Download this file and store it in the directory you will work in. The code below contains calls to these functions.

In this exercice, we will investigate the nucleosome organization around CTCF and ETS1 binding sites in GM12878 cells. CTCF is a factor which is known to show nicely phased nucleosomes arrays on each side of its binding sites. ETS1 is a transcription factor which is mostly found near transcription start sites (TSS) but which does not behave as CTCF regarding nucleosome organization.
We will first extract CTCF and ETS1 peaks from peak lists which have the corresponding factor motif in their neighbourhood using FindM. Then, we will measure the density of MNase and DNAseI reads around these peaks by generating matrices of tag counts using ChIP-Extract. We will import these data in R and perform a probabilistic partitioning to try to identify several nucleosomes organization classes around transcription factor binding sites and isolate classes of interest. Finally, we will extract one class of interest and go back to ChIP-Cor for additional analyses.

  1. First, let's extract CTCF and ETS1 peaks which have CTCF and ETS1 binding motif nearby, respectively. To do this, go the FindM and fill the parameters as follows:

    click to show/hide FindM parameters.

    From the result page you will need to do two things. First save the results on your hard drive under sga format and name the files 'findm_CTCF.sga' and 'findm_ETS1.sga' (in case you can retrieve these files here : findm_CTCF.sga findm_ETS1.sga). Second, redirect the results toward ChIP-Extract by clicking on the corresponding button under 'Useful links for Downstream Analysis'.

  2. For each FindM result, run ChIP-extracts (4 runs total) as shown under and save ChIP-Extract results on your hard drive by naming your files 'chipextract_CTCF_MNase.mat','chipextract_CTCF_DNaseI.mat', 'chipextract_ETS1_MNase.mat', 'chipextract_ETS1_DNaseI.mat'.

    click to show/hide ChIP-Extract parameters.
  3. Now open a R session in the folder where you have saved all the files and load the ChIP-Extract data.
    click to show/hide R code.

  4. To verify whether CTCF has nicely phased nucleosomes arrays on each sides of its binding sites, and that ETS1 does not, you can draw aggregation plots. An aggregation plot is the type of plot that ChIP-cor returns you. ChIP-Extract data can be used to draw exactly the same plots as you would have had with ChIP-Cor.
    click to show/hide R code.

  5. Now, let's run the EM partitioning to classify MNase signal and highlight whether different types of nucleosome organization are present in the data. We will allow the algorithm to shift and flip the data.
    Nucleosome signal is perdiodic by nature. Allowing the vectors to be phased by shifting them will increase the resolution of the signal. For instance, if the nucleosomes are not always organized the same around the binding site, this can create interference and mess up the nucleosome signal. Hopefully, this can be fixed by allowing shifting. Additionally, allowing algorithm to flip the vector might help resolving symetrical patterns which may originate from the overlap of two asymetrical anti-sense patterns.
    click to show/hide R code.

    click to show/hide the figures.

    The result figures shows a few interesting things.
    Regarding CTCF, a strong periodic MNase signal is still visible. Some classes are asymetric, suggesting that the CTCF sites may have an orientation. Moreover, the different classes show pretty different patterns, supporting different type of nucleosome organization around CTCF binding sites. Interestingly, the DNaseI footprint is still visible which indicates that the data were not shifted too much. This suggests that the nucleosomes are organized with respect to the binding sites (or a feature really close).
    Regarding ETS1, we now see nucleosomes clearly visible. However, in contrast to CTCF, DNaseI footprint dissapeared and the DNaseI hypersentitivty region is now wider. This reflects the fact that the MNase data vectors have been shifted. This means that the first nucleosome is not a constant distance from the binding site. This suggests that the nucleosomes are not organized with respect to the binding sites. Additionally, unlike what the aggregation plot was showing, the nucleosome organization around ETS1 binding sites is not anymore symetrical. For instance, class 1 and 3 have a nucleosome array on one side of the binding site only.

  6. This algorithm performs a probabilistic (or soft) classification. This means that each binding site is assigned to all classes at the same time, with different probabilities. Nonetheless, it is still possible to extract all binding sites which have been assigned to a class, with a given minimal probability. Interestingly, it is also possible to update the original binding site position to take into account the shift and flip performed by the algorithm. In case, you can download the classification result file here : CTCF_5class_flip_15shift_random.rds ETS1_5class_flip_15shift_random.rds
    We know that ETS1 is a factor that mostly bind near TSS. Class 1 shows a profile which is compatible with this. The nucleosome array might reflect the nucleosomes downstream the TSS with the commonly described nucleosome depleted region near the TSS. Let's extract the binding sites which have been classified has belonging to class 1 with an overall probability of 0.9 or higher.
    click to show/hide R code.

  7. Now, go on ChIP-Cor, upload the coordinates extracted from class 1 and run ChIP-cor 3x as indicated under and save the results in the files named as follows : 'findm_ETS1_EM_class1_TSS.dat', 'findm_ETS1_EM_class1_MNase.dat', 'findm_ETS1_EM_class1_H3K4me3.dat'

    click to show/hide ChIP-Cor parameters.
  8. Go back to your R session and draw an aggregation plots with all the ChIP-Cor data.
    click to show/hide R code.

    click to show/hide the figure.

    The peak of TSS density in the nucleosome depleted region together with the H3K4me3 labelling suggest that the binding sites which have been assigned to this class are indeed in promoters regions. However, the exact orientation of these promoters is hard to estimate since H3K4me3 labelling is pretty symetrical and that we can see nucleosomes on both sides of the TSS density peak.
    Now, you might want to pursue this analysis by extracting other classes coordinates or to rerun an entire analysis on other type of data such as histone marks.

Part IV: H3K4me3 around human promoters. Dataset in this example is from GEO series GSE29611 and EPDNew.