EpiCODE

epicode.py - Discovers "epigenetic codes" within ChIP-seq datasets.

$ epicode.py absolute -bed [BED6+ file] -bams [BAM files] [options] $ epicode.py differential -bed [BED6+ file] -abams [BAM files] -bbams [BAM files] [options] $ epicode.py discriminatory -beds [BED6+ files] -bams [BAM files] [options]

To get help specific to these three methods see:

$ epicode.py {absolute, differential, discriminatory} --help

The goal of epicode is to discover patterns of histone modifications from aligned sequence data. Epicode looks for combinations (subsets) of marks that tend to occur in (at least) sub-portions of the data. Alternatively it identifies combinations of marks that change coordinately i.e. are "gained" or "lost" frequently at the same time.

The algorithm provides three modes absolute, discriminatory, and differential. The first two modes identify co-occurring marks within one or many sets of genomic losi, respectively. The differential mode attempts to find patterns of coordinated mark changes. In discriminatory mode two (or more) genomic loci are differentiated based their associated patterns.

Walk-through example

To see how epicode could be used with ENCODE data see the full example at: data/encode_3step.md

EpiCODE modes

Each of the provided modes corresponds to a specific subcommand of epicode.

• absolute for experiments with multiple histone modifications or epigenetics marks mapped in a single condition. Epicode finds patterns (epigenetic codes) of frequently co-occurring marks.

• differential for experiments with the same marks mapped in two conditions. Epicode finds patterns of coordinated mark changes i.e. subsets of marks that are often "gained" or "lost" at the same time.

• discriminatory for experiments where one is interested in the epigenetic patterns that distinguish different sets of (preferably non-overlapping) genomic loci. Multiple histone modifications are mapped in a single condition and quantified for two two or more (experimental) sets of loci.

As input epicode expects at least one BED6+ file of reference genomic regions (-bed or -beds) and at least one set of aligned sequence reads in coordinate sorted BAM files. Epicode is not filtering duplicate reads, please run samtools dedup to create deduplicated input files if this is desired.

EpiCODE tasks

The three provided high-level modes are wrappers around tasks with more fine-grained options. A list of all the available tasks can be seen by:

$ epicode.py --help ... | extract_absolute Processes multiple bam files in "absolute" mode. | | discriminatory Discriminatory "epigenetic codes" that emphasize | epigenetic differences between two (or more | [experimental]) sets of sites. | | scale_features Scales features of any input array (loci x features) | using any of the supported algorithms. | | extract_diff Processes multiple bam files in "differential" mode. | | differential Differential "epigenetic codes" from "gain-loss" | changes in levels of epigenetic marks from two | experimental conditions in single set of sites. | | scale_pairs Scales observed paired columns of read-overlap counts. | | code_sklearn Non-negative matrix factorization using scikits-learn. | | multi_code_sklearn Multi-array Non-negative matrix factorization using | scikits-learn. | | scale_diff Calculates differential features from paired counts | array (paired loci x features). | | absolute Absolute "epigenetic codes" from levels of epigenetic | marks in a single experimental condition and in a | single set of sites. ...

High-level interface

Absolute mode: absolute

Designed to work on epigenetic marks mapped in one condition and quantified within one type of loci. The sites should be provided as a BED6+ file e.g. a promoter file or an enhancer file. The sequencing data should be provided as coordinate sorted bam files (e.g. using samtools or novosort). The algorithm results (files) are saved into the odn directory (default absolute_out and prefixed with runid, which defaults to an automatically generated and likely unique integer. Column names within all output files that are data matrices are generated from input BAM filenames and are optionally shortened (--shorten option) by removing redundant substrings.

This is a wrapper for the following chain of tasks, each task saves the generated data as intermediate files in a sigle run directory:

1. extract_absolute - Extracts counts of reads overlapping genomic regions and normalizes by region length.
2. scale_features - Scales features (columns) within each lvl.arr array.
3. code_sklearn - Learn discriminatory epigenetic codes.

The procedure creates two files for the two matrices {parameters}.arr (optionally) and {parameters}.epi.

Example command line:

epicode.py absolute -c 6 -par 4 -bed <<repo dir>>/data/h1esc_prom_5000.bed -bams <<bam dir>>/*.bam

Here <<repo dir>> is where you checked out the git reposity and <<bam dir>> is a directory with bam files. The parameters -c 6 and -par 4 mean that six codes in "absolute" mode will be learned and the bam file processing will happen using four cores.

Differential mode: differential

Designed to work on epigenetic marks mapped in two conditions (A and B) quantified in one type of locus. The sites should be provided as a BED6+ file e.g. a promoter file or an enhancer file. The sequencing data should be provided as coordinate sorted bam files (e.g. using samtools or novosort). The algorithm results (files) are saved into the odn directory (default differential_out and prefixed with runid, which defaults to an automatically generated and likely unique integer. Column names within all output files that are data matrices are generated from input BAM filenames and are optionally shortened (--shorten option) by removing redundant substrings.

The input to the differential mode are two ordered sets of BAM files. Each set should contain several BAM files, typically more than 4. Within each ordered set the bam files should have identical names i.e. they should be the same marks mapped in two conditions.

This is a wrapper for the following chain of tasks, each task saves the generated data as intermediate files in a sigle run directory:

1. extract_differential - Extracts paired counts within genomic regions in step resultion.
2. scale_pairs - Normalizes counts paired samples for sequencing depth.
3. scale_differential - Converts scaled absolute counts to gain-loss levels.
4. scale_features - Scales features (columns) within each {parameters}lvl.arr array.
5. code_sklearn - Learn absolute epigenetic codes.

The procedure creates two files for the two matrices {parameters}.arr (optionally) and {parameters}.epi.

Example command line:

epicode.py differential -c 8 -par 4 <<repo dir>>data/hsmm_prom_1000.bed -abams <<A bams dir>>/*.bam -bbams <<B bams dir>>/*.bam Here <<repo dir>> is where you checked out the git reposity and <<A bams dir>> and <<B bams dir>> are directories with BAM files. Epicode assumes that both directories contain BAM files with identical names, such that listing their contents returns identical filenames (basenames). The parameters -c 6 and -par 4 mean that six codes in "differential" mode will be learned and the BAM file processing will happen using four cores.

Discriminatory mode: discriminatory

Designed to work on epigenetic marks mapped in a single condition, quantified within two (or more) types of loci. The types of sites are provided as BED6+ files e.g. a promoter file and an enhancer file. The sequencing data should be provided as coordinate sorted bam files (e.g. using samtools or novosort). The algorithm results (files) are saved into the odn directory and prefixed with runid, which defaults to discriminatory. Column names within all output files that are data matrices are generated from input BAM filenames and are optionally shortened (--shorten option) by removing redundant substrings.

The algorithm takes three important parameters c the number of expected histone codes and also rank of the factored matrices, colsca the algorithm used to scale the levels (columns) of the final input matrices to the NMF algorithm, and init the algorithm used to initialize matrices. The two latter parameters are best left as defaults. The c parameter has no default as it depends both on the number of input bam files, redundancy (correlation) of the assayed epigenetic marks and the biological complexity of the genomic regions (bed files). Typically a value between 4-10 gives interpretable results, but please see our publication for some recommendations and properties of NMF applied to epigenomic data.

This is a wrapper for the following chain of tasks, each task saves the generated data as intermediate files in a sigle runid directory:

1. extract_absolute - Extracts mark lavels *lvl.arr for all marks (BAM files) and regions (BED files) combinations.
2. scale_features - Scales features (columns) within each *lvl.arr array.
3. multicodesklearn - Learn discriminatory epigenetic codes.

The procedure creates two files for the two matrices {parameters}.arr (optionally) and {parameters}.epi. The files can be used as input to the logistic regression classifier task logistic_classifier.

epicode.py discriminatory -par 4 -beds <<repo dir>>data/pol2_1000_e25_*.bed -c 7 -bams <<bam dir>>/*.bam

Here <<repo dir>> is where you checked out the git reposity and <<bam dir>> is a directory with bam files. The parameters -c 7 and -par 4 mean that seven codes in "differential" mode will be learned and the bam file processing will happen using four cores.

Low-level interface

code_sklearn

NMF factorization of asingle array. Creates two new files with .epi and .arr suffixes (only with --transform) that include algorithm parameters in their names in the same directory as the input. Method and remaining arguments are currently ignored from the command line. (see: sklearn.decomposition.NMF).

multi_code_sklearn

NMF factorization of multiple arrays. Creates two new files with .epi and .arr suffixes (only with --transform) that include algorithm parameters in their names in the same directory as the input. Method and remaining arguments are currently ignored from the command line. (see: sklearn.decomposition.NMF).

scale_features

Scales features of any input array (loci x features) using any of the supported algorithms. Creates new array file with scaled columns scaling algorithm name is included as suffix. See: scarrfunction documentation for details.

scale_pairs

Scales observed paired columns of read-overlap counts (paired samples). Currently only the deseq algorithm is implemented.

scale_diff

Calculates differential features from paired counts array (paired loci x features). A proper input for this function is obtained by the extract_differential sub-command. Matches pairs by an :a and :b suffix. Creates a new file with _lvl.arr suffix in the same directory as the input. The gain and loss columns are suffixed :g and :l, respectievely.

extract_absolute

It estimates enrichment levels for each mark at each input genomic regions from the BED6+ file. Read count extraction can be done in parallel (per-chromosome parallelism). The output is saved into the odn directory with runid prefix (optional) and has an extension of {runid}_lvl.arr. If shorten is enabled \ --shortenthe algorithm will try to shorten the file names when producing column names by removing common substrings. In the case of erros see the log messages. Program aborts if output files are present.

Example command-line:

epicode.py extract_absolute -bed test.bed -bams directory_with_bams/*bam -odn test_extract -par 8

extract_differential

For each bam file it counts the number reads overlapping each genomic region from the BED6+ file. Read count extraction can be done in parallel (per-chromosome parallelism). The output is saved into the odn directory with runid prefix (optional) and has an extension of {runid}_cnt.arr. If shorten is enabled --shorten the algorithm will try to shorten the file names when producing column names by removing common substrings. Column names receive a :a or :b suffix when they are from tha A-list or B-list, respectively. In the case of erros see the log messages. Program aborts if output files are present. Output array is proper input for the scale_features task.

Plotting

Two scripts are provided to facilitate plotting. They can be found in the epicode/scripts directory. Each script take an epigenetic code file .epi and the name of an output graphics file supported by ggplot e.g. png, pdf etc.

Example usage:

$ Rscript <<repo dir>>/scripts/absolute_plot.r epicode/data/absolute_codes.epi absolute_codes.png

$ Rscript <<repo dir>>/scripts/gainloss_plot.r epicode/data/differential_codes.epi differential_codes.png

The above scripts require the following packages: ggplot2, reshape2, stringr. All of these can be installed from CRAN by the following command:

R install.packages("<<package name>>")

For example:

R install.packages("ggplot2")

Logging Configuration

Epicode.py can be configured to use an external configuration file and to log to an arbitrary file stream

See moke help for details:

$ epicode.py --help

Installation

Automatic Installation

In the simple case installing Epicode requires only:

bash $ easy_install-2.7 epicode

If the above command is not found first try:

$ easy_install epicode

If the installation procedure runs, but fails at some point follow the manual installation guide. If easy_install is still not found please try installing setuptools first (see below).

Manual Installation

Since the automatic installation procedure failed we have to make sure that we are running the correct version of Python and easy_install (setuptools):

Python and Setuptools

$ python2 --version | Python 2.7.5

Verify that easy_install can be found:

$ which easy_install | .../bin/easy_install

If you cannot find easy_install you have to install setuptools this is best done system-wide, using the specific mechanisms. We will use easy_install to install additional dependencies later on.

Arch:

$ pacman -S python2-setuptools

Fedora:

$ yum install python-setuptools python-setuptools-devel

Ubuntu:

$ sudo apt-get install python-setuptools python-dev

Alternatively one can try to follow the instructions at:

https://pypi.python.org/pypi/setuptools/0.6c11

Additional Dependencies

Epicode has a small number of dependencies. On many systems they will be successfully installed using the setuptools/easy_install mechanism. However, we recommend to install numpy and scipy using the sytem-wide mechanism. As their compilation is particularly involved. Epicode was tested and developed for numpy-1.7.1 and scipy-0.12.0, but should also work on other relatively recent releases.

For Arch linux:

$ pacman -S extra/python2-numpy community/python2-scipy

Fedora:

$ yum install numpy scipy

Ubuntu:

$ sudo apt-get install python-numpy python-scipy

For other operating systems follow the instructions at:

http://www.scipy.org/install.html

Next, we will install pysam, scikit-learn, and moke from PyPI:

$ easy_install pysam==0.7.5 scikit-learn==0.14.1 moke==1.1.5

If all the commands returned correctly you should be able to start python:

$ python2

And issue the following statements:

```python

import numpy import scipy import sklearn import moke import pysam ```

Now you are ready to install epicode

$ easy_install epicode