darked89 / MUSIC

MUltiScale enrIchment Calling for ChIP-Seq Datasets

<title>MUSIC</title>

MUSIC is an algorithm for identification of enriched regions at multiple scales in the read depth signals from ChIP-Seq experiments.
It takes as input:

- Mapped ChIP and control reads (optional),
- Smoothing scale window lengths (in base pairs),

and outputs:

- Enriched regions at multiple scales,
- Significantly enriched regions from all the scales.

Unlike other ER identification methods, MUSIC allows analyzing the scale length spectrum of ChIP-Seq datasets and also selecting a user specific slice in the scale length spectrum with custom granularity and generates the enriched regions at each length scale.

MUSIC does not strictly require a control experiment (for example mock IP, input DNA) to be performed. It is, however, strongly advised to generate control datasets matching the sample of interest (See Landt et al 2012).

Download and Installation

You can download MUSIC C++ code here. There are no dependencies for building MUSIC. After download, type:

unzip MUSIC.zip
cd MUSIC
make clean
make

to build MUSIC. The executable is located under directory bin/. It may be useful to install samtools for processing BAM files.

To get help on which options are available, use:

MUSIC -help

Usage

MUSIC run starts with preprocessing the reads for ChIP and control samples (Note that we use samtools for converting BAM file to SAM files.):

mkdir chip;mkdir input
samtools view chip.bam | MUSIC -preprocess SAM stdin chip/
samtools view input.bam | MUSIC -preprocess SAM stdin input/

If there are multiple replicates to be pooled, they can be done at once or separately. If done separately, MUSIC pools the reads automatically. Then it is necessary to sort and remove duplicate reads in control and ChIP samples:

mkdir chip/sorted;mkdir chip/dedup;mkdir input/sorted;mkdir input/dedup
MUSIC -sort_reads chip chip/sorted
MUSIC -sort_reads input input/sorted
MUSIC -remove_duplicates chip/sorted 2 chip/dedup
MUSIC -remove_duplicates input/sorted 2 input/dedup

We do enriched region identification:

MUSIC -get_multiscale_broad_ERs -chip chip/dedup -control input/dedup -mapp Mappability_36bp -l_mapp 36 -begin_l 1000 -end_l 16000 -step 1.5

This code tells MUSIC to identify the enriched regions starting from 1kb smoothing window length upto 16kb with multiplicative factor of 1.5 using the default parameters for the remaining parameters. The ERs for each scale are dumped.

In case there is no control, skip the option that specifies preprocessed control reads directory (i.e., '-control input/dedup')

There are 3 different ER identification modes by default.

-get_TF_peaks

Identifies point binding events. This uses very small scale level to identify the small transcription factor binding peaks. Use this mode for TF's like CTCF. MUSIC aims at trimming the reported regions and identifying the peaks with most dense signal in it.

-get_multiscale_punctate_ERs

Identifies punctate enriched regions. Uses 100 base pairs to 2 kbps scale levels. This mode is useful for punctate histone marks like H3K4me3, H3K27ac, and marks that behave in a mixed manner with a dominating spectrum at punctate scales like H3K4me1.

-get_multiscale_broad_ERs

Identifies the ERs at broad scales that can span megabases. Use this option for marks like H3K9me3, H3K27me3, H3K36me3, H3K79me2, H4K20me1... Make sure you select p-value normalization parameter to balance power and false positive rate.

MUSIC can save the smoothed tracks in bedGraph format that can be viewed locally:

MUSIC -write_MS_decomposition -chip chip/dedup -control input/dedup -mapp Mappability_36bp -l_mapp 36

The smoothed bedGraph files are usually very small in size and can easily be stored/transferred.

Output format

MUSIC output a large number of files that contain the SSERs at different scales (named SSER_....bed).

The final set of ERs are reported in two files: One is broadPeak formatted (http://genome.ucsc.edu/FAQ/FAQformat.html#format13)

Other file is in an extended BED format and has 9 columns:

[chromosome] [start] [end] ["."] [log_10 Q-value] [Strand ("+")] [Summit position] [Mappable Trough Position] [Fold Change]

The entries are sorted with respect to increasing Q-values.

Note that MUSIC reports log_10(Q-values) and not -log_10(Q-value).

Issue with Vieweing BED file in IGV

If you are vieweing the BED file with IGV, it sometimes does not plot the ERs correctly. If you wish to look at the bed file in IGV, use only the first 6 columns in the file:

cut -f1-6 ERs_1000.0_16000.0_1.50_1750_4.0.bed > ERs.bed

then open ERs.bed in IGV.

Parameter Selection Guideline for the Studied ChIP-Seq Datasets

MUSIC has a set of default parameter sets for broad marks (like H3K36me3, H3K27me3), punctate marks (H3K4me3), and point binding (like transcription factors) which work well in most cases. When one is not sure about the ER scale spectrum for a signal profile at hand, one can perform a scale spectrum analysis using a large spectrum with dense sampling and get an idea about the dominant ER length scale (if there is one) for the signal profile and match the parameters used in the manuscript.

Parametrization for New ChIP-Seq Datasets

When a new ChIP-Seq dataset is going to be processed, it is necessary to choose begin and end length scales and p-value normalization window length. The length scales enables one to concentrate on the correct scale spectrum and p-value normalization window length compensates for variation in sequencing depth and allows controlling the estimated false positive rates.

There are two steps to parameter selection:

1. Select a stringent p-value normalization window length: -get_per_win_p_vals_FC option.

This option estimates the false positive and negative rates using a large selection of p-value window lengths. The output is a file where each row look like this:
...
l_win: 1700 FNR: (FC:0.001) (p-val:0.001) FPR: (FC:0.010) (p-val:0.005) Sentitivity: 0.999
...

This option evaluates several windows lengths and estimates the false positive rate and false negative rates. We recommends using the maximum window length (l_win) where false positive rate (FPR) for FC and p-val values are below 1%.

Note that you can skip going through the file manually if you specify p-value normalization window length as 0 ('-l_p 0') in the peak calling step; which tells MUSIC to select p-value normalization window length from the above file automatically. For this to work, make sure that you do not delete any files after running -get_per_win_p_vals_FC option, otherwise MUSIC will complain that it cannot find the file. See below for complete automation (of p-value normalization window length selection) with default parameters.

2. Using the p-value normalization window length in step 1, generate the scale specific ER scale spectrum: -get_scale_spectrum option.

MUSIC generates the spectrum (using the scale lengths 100 base pairs to 1 megabase. The output is reported in a text file where each row corresponds to a scale length. In each row, the coverage of the SSERs is given. For example:
...
27 56815.13 111 11675769
...

where 2nd column is the scale length and 4th column is the coverage of the ERs that are specific to that scale. It is best to plot the spectrum, i.e., the scale lengths versus the fraction of coverage of the SSERs (4th column in the file), then match the spectrum with the studied HMs in the manuscript. If the spectrum is very different from all of the parametrized ChIP-Seq datasets, it is useful to generate the statistics on ER length distribution and distribution of ER-ER distances and use the procedure described in the manuscript to select the scale levels.

The other parameters (namely, gamma and sigma) do not depend on experimental variables and are optimized for minimizing overmerging and maximizing sensitivity. We suggest using the values specified in the manuscript, i.e., gamma=4, sigma=1.5.

Important Note on Punctate ERs

After this analysis, if the scale spectrum turns out to be punctate, i.e., the dominating scale is smaller than 10kb; there is a possibility that the p-value normalization window length parameter (l_p) yields low sensitivity. To compensate for this, we recommend running with default l_p parameter of MUSIC.

Running MUSIC with Default Parameters and Automatic Selection of l_p Parameter:

We just added a new script run_MUSIC.csh. This script automates the parameter selection for -l_p option. This script automates running MUSIC with default parameters. It simply calls MUSIC with above parameters in order. Here is how this script can be used:

read_fp="wgEncodeBroadHistoneGm12878H3k27me3StdAlnRep1.bam";
mappability_map="mappability/36bp";
input_processed_dir="input/pruned";
mkdir preprocessed sorted pruned
run_MUSIC.csh -preprocess ${read_fp} preprocessed
run_MUSIC.csh -remove_duplicates preprocessed sorted pruned
run_MUSIC.csh -get_optimal_broad_ERs pruned ${input_processed_dir} ${mappability_map}

Make sure that run_MUSIC.csh is included in PATH. Currently only BAM files are supported in preprocessing. We will add more file types, soon. Note that it is necessary to install samtools for making sure this runs smoothly. The above code calls the ER's using the optimal l_p selection with the default parameters.

Multi-Mappability Signals

Using Mappability correction increases the accuracy of MUSIC. You can download the multi-mappability signals for several common read lengths here.

Multi-Mappability Profile Generation

To generate the multi-mappability profile, MUSIC depends on a short read aligner. By default, bowtie2 generated SAM alignments are supported. Following are necessary for multi-Mappability profile generation:

- The FASTA file for the genomic sequence with all the chromosomes
- The read length
- bowtie2 installation and the genome indices for bowtie2
- Read length

Use this script generate_multimappability_signal.csh under bin/ directory to generate the multi-mappability profile. For example, to generate the 50 bp multi-Mappability profile for human genome,

cd bin
./bin/generate_multimappability_signal.csh hg19.fa 50 /home/users/music_user/bt2_indices/hg19_indexes/hg19

Note that the bowtie indices must be supplied in the command line and the path must be the absolute path. For the above command, they are under /home/users/music_user/bt2_indices/hg19_indexes/hg19.

This command processes the FASTA file and writes two temporary scripts (temp_map_reads.csh, temp_process_mapping.csh) that fragments the genome, maps the fragments, and then generates the profile. Next, it is necessary to run these two scripts, in order:

./temp_map_reads.csh
./temp_process_mapping.csh

After the scripts are run, multi-mappability profile for each chromosome should be created with '.bin' extension. 'temp_map_reads.csh' is the most time consuming script that maps the fragments to the genome. Each line in the script is a command that maps the reads to a chromosome that can be run in parallel on a cluster. It is important to make sure 'temp_map_reads.csh' finishes before running 'temp_process_mapping.csh'.

Your can also email me ([e-mail]) to generate multimappability profiles for new species.

Datasets

The ENCODE datasets can be downloaded from UCSC Genome Browser.
The H3K36me3 datasets for K562 and GM12878 cell lines can be downloaded from here.

Arif Harmanci ([e-mail]), Joel Rozowsky, Mark Gerstein, 2014