hicpeaks 0.3.4

Identify real loops from Hi-C data.

HiCPeaks

hicpeaks provide a Python CPU-based implementation for BH-FDR and HiCCUPS, two peak calling algorithms for Hi-C data, proposed by Rao et al [1].

Installation

hicpeaks is developed and tested on UNIX-like operating system, and following packages or softwares are required:

Python requirements:

1. Python 2.7/3.5+

2. Multiprocess

3. Numpy

4. Scipy

5. Matplotlib

6. Pandas

7. Statsmodels

8. Scikit-Learn

9. H5py

10. Cooler

Other requirements:

• ucsc-fetchchromsizes

conda, an excellent package manager, can be used to install all requirements above.

Install Conda

Note

If you have the Anaconda Distribution installed, you already have it.

Choose an appropriate Miniconda installer for your system, then in your terminal window type the following and follow the prompts on the installer screens:

$ bash Miniconda2-latest-Linux-x86_64.sh

After that, update the environment variables to finish the Conda installation:

$ source ~/.bashrc

Install Packages through Conda

First set up the channels to make all packages listed above accessible (note that the order is important to guarantee the correct priority):

$ conda config --add channels defaults
$ conda config --add channels bioconda
$ conda config --add channels conda-forge

Then just type and execute the following command:

$ conda install numpy scipy matplotlib pandas statsmodels scikit-learn h5py multiprocess cooler ucsc-fetchchromsizes

Install hicpeaks

Now download the hicpeaks source code from PyPI, extract it and run the setup.py script:

$ python setup.py install

Overview

hicpeaks comes with 6 scripts: toCooler, pyBHFDR, pyHICCUPS, combine-resolutions, peak-plot and apa-analysis.

• toCooler

  Store TXT/NPZ bin-level Hi-C data into cooler container.

  1. I have included a sample data with hicpeaks source code to illustrate how you should prepare your data in TXT format. It’s quite easy, just remember 3 points: 1. the file name should follow this pattern “chrom1_chrom2.txt” (remove prefix from your chromosome labels, i.e. “chr1” should be “1”, and “chrX” should be “X”); 2. each file should only contain 3 columns, corresponding to “bin1” of “chrom1”, “bin2” of “chrom2”, and “contact frequency” (don’t perform any normalization processes); 3. all files at the same resolution should be placed under a single folder.

  2. NPZ format is another bin-level Hi-C data container which can extremely speed up data loading. hicpeaks supports NPZ files generated by old version of runHiC (<0.8.0) and TADLib (<0.4.0).

• pyBHFDR

  A CPU-based python implementation for BH-FDR algorithm. Rao et al (2014) states in their supplementary material that this algorithm is robust enough to obtain all main results of their paper. Compared with HiCCUPS, BH-FDR doesn’t use λ-chunk in multiple hypothesis test, and only considers the Donut background region when calculating the expected values.

• pyHICCUPS

  A CPU-based python implementation for HiCCUPS algorithm. Besides the donut region, HiCCUPS also considers the lower-left, vertical and horizontal backgrounds when calculating the expected values. And λ-chunk is used to overcome several multiple hypothesis testing challenges for Hi-C data. Finally, while BH-FDR has to limit the detected pixels near the diagonal (<2Mb), HiCCUPS is able to generalize itself to any genomic distance in theory. Here, pyHICCUPS keeps all main concepts of the original algorithm except for these points:

  1. pyHICCUPS excludes vertical and horizontal backgrounds from its calculation.

  2. There are two critical parameters related to the loop definition in HiCCUPS: the peak width p and the donut width w. In original implementation, they are set exclusively for each certain resolution, specifically, p=1 and w=3 at 25Kb, p=2 and w=5 at 10Kb, and p=4 and w=7 at 5Kb. To improve the sensitivity, pyHICCUPS calculates and outputs the union of the peak calls from all parameter settings (1,3), (2,5), (4,7) in a single run.

  3. Due to computational complexity, you should still limit the genomic distance of 2 loci to some degree (5Mb/10Mb).

• combine-resolutions

  Combine peak calls from different resolutions in a way similar to original HiCCUPS. Briefly, it excludes redundant lower resolution peaks while filters out low-confidence high resolution peaks.

• peak-plot

  Visualize peaks (or loops) detected by pyBHFDR or pyHICCUPS on heatmap. Just provide a cooler file, a loop annotation file in bedpe format, and input your interested region (chrom, start, end).

• apa-analysis

  Perform Aggregate Peak Analysis (APA). The inputs are a Hi-C matrix in .cool format and a loop annotation file in .bedpe format.

QuickStart

This tutorial will guide you through the basic usage of all scripts distributed with hicpeaks.

toCooler

If you have already created a cooler file for your Hi-C data, skip to the next section pyBHFDR and pyHICCUPS, go on otherwise.

First, you should store your TXT/NPZ bin-level Hi-C data into a cooler file by using toCooler. Let’s begin with our sample data below. Suppose you are still in the hicpeaks distribution root folder: change your current working directory to the sub-folder example:

$ cd example
$ ls -lh *

-rw-r--r-- 1 xtwang  18 May  4 18:00 datasets
-rw-r--r-- 1 xtwang 293 May  4 18:00 hg38.chromsizes

25K:
total 12M
-rw-r--r-- 1 xtwang 12M May  4 18:00 21_21.txt

There is one sub-directory called 25K which contains interactions within the smallest chromosome in K562 cell line at 25K resolution, and one metadata file datasets which we can pass directly to toCooler:

$ cd 25K
$ head -5 21_21.txt

201 703     1
201 1347    1
201 1351    1
201 1524    1
201 1691    1

$ cd ..
$ cat datasets

res:25000
  ./25K

You should construct your TXT files (no head, no tail) with 3 columns, which indicate “bin1 of the 1st chromosome”, “bin2 of the 2nd chromosome” and “contact frequency” respectively. See Overview above.

To transform this data to cooler format, just run the command below:

$ toCooler -O K562-MboI-parts.cool -d datasets --assembly hg38 --nproc 1

toCooler routinely fetch sizes of each chromosome from UCSC with the provided genome assembly name (here hg38). However, if your reference genome is not holded in UCSC, you can also build a file like “hg38.chromsizes” in current working directory, and pass the file path to the argument “–chromsizes-file”.

Type toCooler with no arguments on your terminal to print detailed help information for each parameter.

For this datasets, toCooler will create a cooler file named “K562-MboI-parts.cool”, and your data will be stored under the URI “K562-MboI-parts.cool::25000”.

This tutorial only illustrates a very simple case, in fact the metadata file may contain list of resolutions (if you have data at different resolutions for the same cell line) and corresponding folder paths (both relative and absolute path are accepted, and if your data are NPZ format, this path should point to the NPZ file):

res:10000
  /absoultepath/10K

res:25000
  ../relativepath/25K

res:40000
  /npzfile/anyprefix.npz

Then toCooler will generate a single cooler file storing all the specified data under different cooler URI: “specified_cooler_path::10000”, “specified_cooler_path::25000” and “specified_cooler_path::40000”.

pyBHFDR and pyHICCUPS

With cooler URI, you can perform peak annotation by pyBHFDR or pyHICCUPS:

$ pyBHFDR -O K562-MboI-BHFDR-loops.txt -p K562-MboI-parts.cool::25000 -C 21 --pw 1 --ww 3

Or:

$ pyHICCUPS -O K562-MboI-HICCUPS-loops.txt -p K562-MboI-parts.cool::25000 --pw 1 2 4 --ww 3 5 7 --only-anchors

Type pyBHFDR or pyHICCUPS on your terminal to print detailed help information for each parameter.

Before step to the next section, let’s list the contents under current working directory again:

$ ls -lh

total 852K
drwxr-xr-x 4 xtwang  128 May  4 18:21 25K/
-rw-r--r-- 1 xtwang  17K May  4 18:23 K562-MboI-BHFDR-loops.txt
-rw-r--r-- 1 xtwang  15K May  4 18:23 K562-MboI-HICCUPS-loops.txt
-rw-r--r-- 1 xtwang 723K May  4 18:22 K562-MboI-parts.cool
-rw-r--r-- 1 xtwang   18 May  4 18:21 datasets
-rw-r--r-- 1 xtwang  293 May  4 18:21 hg38.chromsizes
-rw-r--r-- 1 xtwang 2.2K May  4 18:23 pyBHFDR.log
-rw-r--r-- 1 xtwang 8.5K May  4 18:23 pyHICCUPS.log
-rw-r--r-- 1 xtwang  17K May  4 18:22 tocooler.log

The detected loops are reported in a customized bedpe format. The first 10 columns are identical to the official definition, and the additional fields are:

11. Fold enrichment score calculated from the donut background.

12. The p value calculated from the donut background.

13. The q value calculated from the donut background.

14. Fold enrichment score calculated from the lower-left background.

15. The p value calculated from the lower-left background.

16. The q value calculated from the lower-left background.

Peak Visualization

Now, you can visualize BH-FDR and HICCUPS peak annotations on heatmap with peak-plot.

For HICCUPS peaks:

$ peak-plot -O test-HICCUPS.png --dpi 200 -p K562-MboI-parts.cool::25000 -I K562-MboI-HICCUPS-loops.txt -C 21 -S 25000000 -E 31000000 --correct

The output figure should look like this:

Aggregate Peak Analysis

To inspect the overall loop patterns of the detected peaks, you can use the apa-analysis script:

$ apa-analysis -O apa.png -p K562-MboI-parts.cool::25000 -I K562-MboI-HICCUPS-loops.txt -U

The output plot should look like this:

Combine different resolutions

The inputs to combine-resolutions are loop annotation files (bedpe) at different resolutions. If an interaction is detected as a peak in both resolutions, this script records the precise coordinates in finer resolutions and discards the coarser resolution one. And a long-range (determined by the --min-dis parameter) peak call at high resolutions (determined by the --good-res parameter) will be treated as a false positive if it could not be identified at lower resolutions. Here’s a pseudo command with 3 loop files at 5Kb, 10Kb, and 20Kb respectively:

$ combine-resolutions -O K562-MboI-pyHICCUPS-combined.bedpe -p K562-MboI-pyHICCUPS-5K.txt K562-MboI-pyHICCUPS-10K.txt K562-MboI-pyHICCUPS-20K.txt -R 5000 10000 20000 -G 20000 -M 100000

Performance

The tables below show the performance test of toCooler, pyBHFDR and pyHICCUPS with low (T47D) and high (K562) sequencing data, at low (40K) and high (10K) resolutions.

• Processor: 2.6 GHz Intel Core i7, Memory: 16 GB 2400 MHz DDR4

• Software version: hicpeaks 0.3.0

• At 40Kb resolution, --pw and --ww are set to 1 and 3 respectively; at 10Kb resolution, they are set to 2 and 5 respectively.

• The original Hi-C data is stored in TXT

• Number of proccesses assigned: 1

• Valid contacts: total number of non-zero pixels on intra-chromosomal matrices

• Running time format: hr: min: sec

Datasets	Valid contacts	toCooler	pyBHFDR	pyHICCUPS
	Memory Usage	Running time	Memory Usage	Running time	Memory Usage	Running time
T47D (40K)	25,216,875	<600M	0:07:55	<600M	0:01:34	<600M	0:04:17
K562 (40K)	49,088,465	<1.2G	0:21:37	<1.0G	0:01:49	<1.0G	0:03:21
K562 (10K)	139,884,876	<3.0G	1:00:07	<2.0G	0:24:53	<4.0G	1:57:33

Note

Both pyBHFDR and pyHICCUPS support multiple processes (--nproc). If your computer has sufficient memory, the calculation should end within 30 minutes even for high resolutions.

Release Notes

Version 0.3.4 (05/04/2019)

• Improved the local clustering efficiency

• Changed output loop format to bedpe

Version 0.3.3 (03/08/2019)

• Float matrix support in toCooler transformation

• Removed ticklabels in APA plot

Version 0.3.2 (03/03/2019)

1. Supported combination of different resolutions

2. Changed local clustering algorithm

3. Added APA module

4. Compatible with cooler 0.8

5. Old distutils to setuptools

Version 0.3.0 (09/03/2018)

1. Removed horizontal and vertical backgrounds for performance

2. Supported multiple parameters (pw and ww)

3. Supported Python 3

4. Optimized the calculation

5. Code refactoring

6. Fixed bugs when users provide with external .cool files.

Version 0.2.0-r1 (08/26/2018)

1. Speeded up the program by dynamically limiting donut width

2. Added performance table in README.rst

Version 0.2.0 (08/25/2018)

1. Added vertical and horizontal backgrounds

2. Added additional filtering based on dbscan clusters and more stringent q value thresholds

3. Fixed bugs in storing interchromosomal data

Version 0.1.1 (08/24/2018)

1. Lower memory usage and more efficient calculation

Version 0.1.0 (08/22/2018)

1. The first release.

2. Added toCooler and peak-plot.

3. Added multiple process support.

Pre-Release (05/04/2015)

1. Implemented core algorithms of BH-FDR and HICCUPS

Reference

[1]

Rao SS, Huntley MH, Durand NC et al. A 3D Map of the Human Genome at Kilobase Resolution Reveals Principles of Chromatin Looping. Cell, 2014, 159(7):1665-80.