Python package to detect chromatin loops (and other patterns) in Hi-C contact maps.
- Associated publication: https://www.nature.com/articles/s41467-020-19562-7
- Documentation and analyses examples: https://chromosight.readthedocs.io
- scripts used for the analysis presented in the article https://github.com/koszullab/chromosight_analyses_scripts
Stable version with pip:
pip3 install --user chromosightStable version with conda:
conda install -c bioconda -c conda-forge chromosightor, if you want to get the latest development version:
pip3 install --user -e git+https://github.com/koszullab/chromosight.git@master#egg=chromosight
The two main subcommands of chromosight are detect and quantify. For more advanced use, there are two additional subcomands: generate-config and list-kernels. To get the list and description of those subcommands, you can always run:
chromosight --helpPattern detection is done using the detect subcommand. The quantify subcommand is used to compute pattern matching scores for a list of 2D coordinates on a Hi-C matrix. The generate-config subcommand is used to create a new type of pattern that can then be fed to detect using the --custom-kernel option. The list-kernels command is used to view informations about the available patterns.
To get a first look at a chromosight run, you can run chromosight test, which will download a test dataset from the github repository and run chromosight detect on it. You can then have a look at the output files generated.
When running chromosight detect, there are a handful parameters which are especially important:
--min-dist: Minimum genomic distance from which to detect patterns. For loops, this means the smallest loop size accepted (i.e. distance between the two anchors).--max-dist: Maximum genomic distance from which to detect patterns. Increasing also increases runtime and memory use.--pearson: Detection threshold. Decrease to allow a greater number of pattern detected (with potentially more false positives). Setting a very low value may actually reduce the number of detected patterns. This is due to the algorithm which might merge neighbouring patterns.--perc-zero: Proportion of zero pixels allowed in a window for detection. If you have low coverage, increasing this value may improve results.
To detect all chromosome loops with sizes between 2kb and 200kb using 8 parallel threads:
chromosight detect --threads 8 --min-dist 20000 --max-dist 200000 hic_data.cool output_prefix
Pattern exploration and detection
Explore and detect patterns (loops, borders, centromeres, etc.) in Hi-C contact
maps with pattern matching.
Usage:
chromosight detect [--kernel-config=FILE] [--pattern=loops]
[--pearson=auto] [--win-size=auto] [--iterations=auto]
[--win-fmt={json,npy}] [--norm={auto,raw,force}]
[--subsample=no] [--inter] [--tsvd] [--smooth-trend]
[--n-mads=5] [--min-dist=0] [--max-dist=auto]
[--no-plotting] [--min-separation=auto] [--dump=DIR]
[--threads=1] [--perc-zero=auto]
[--perc-undetected=auto] <contact_map> <prefix>
chromosight generate-config [--preset loops] [--click contact_map]
[--norm={auto,raw,norm}] [--win-size=auto] [--n-mads=5]
[--threads=1] <prefix>
chromosight quantify [--inter] [--pattern=loops] [--subsample=no]
[--win-fmt=json] [--kernel-config=FILE] [--norm={auto,raw,norm}]
[--threads=1] [--n-mads=5] [--win-size=auto]
[--perc-undetected=auto] [--perc-zero=auto]
[--no-plotting] [--tsvd] <bed2d> <contact_map> <prefix>
chromosight list-kernels [--long] [--mat] [--name=kernel_name]
chromosight test
detect:
performs pattern detection on a Hi-C contact map via template matching
generate-config:
Generate pre-filled config files to use for detect and quantify.
A config consists of a JSON file describing parameters for the
analysis and path pointing to kernel matrices files. Those matrices
files are tsv files with numeric values as kernel to use for
convolution.
quantify:
Given a list of pairs of positions and a contact map, computes the
correlation coefficients between those positions and the kernel of the
selected pattern.
list-kernels:
Prints information about available kernels.
test:
Download example data and run loop detection on it.
Input Hi-C contact maps should be in cool format. The cool format is an efficient and compact format for Hi-C data based on HDF5. It is maintained by the Mirny lab and documented here: https://open2c.github.io/cooler/
Most other Hi-C data formats (hic, homer, hic-pro), can be converted to cool using hicexplorer's hicConvertFormat or hic2cool. Bedgraph2 format can be converted directly using cooler with the command cooler load -f bg2 <chrom.sizes>:<binsize> in.bg2.gz out.cool. For more informations, see the cooler documentation
For chromosight quantify, the bed2d file is a text file with at least 6 tab-separated columns containing pairs of coordinates. The first 6 columns should be chrom start end chrom start end and have no header. Alternatively, the output text file generated by chromosight detect is also accepted. Instructions to generate a bed2d file from a bed file are given in the documentation.
Three files are generated by chromosight's detect and quantify commands. Their filenames are determined by the value of the <prefix> argument:
prefix.tsv: List of genomic coordinates, bin ids and correlation scores for the pattern identifiedprefix.json: JSON file containing the windows (of the same size as the kernel used) around the patterns from pattern.txtprefix.pdf: Plot showing the pileup (average) window of all detected patterns. Plot generation can be disabled using the--no-plottingoption.
Alternatively, one can set the --win-fmt=npy option to dump windows into a npy file instead of JSON. This format can easily be loaded into a 3D array using numpy's np.load function.
Note: the p-values and q-values provided in prefix.tsv should not be used as a criterion for filtering and are only useful for ranking calls. Their values are obtained from a Pearson correlation test and could be biased due to the dependence between contact values in the window.
All contributions are welcome. We use the numpy standard for docstrings when documenting functions.
The code formatting standard we use is black, with --line-length=79 to follow PEP8 recommendations. We use nose2 as our testing framework. Ideally, new functions should have associated unit tests, placed in the tests folder.
To test the code, you can run:
nose2 -s tests/Questions from previous users are available in the github issues. You can open a new issue for your question if it is not already covered.
When using Chromosight in you research, please cite the pubication: https://www.nature.com/articles/s41467-020-19562-7