pygrgl 2.8

Genotype Representation Graph Library

Genotype Representation Graphs

A Genotype Representation Graph (GRG) is a compact way to store reference-aligned genotype data for large genetic datasets. These datasets are typically stored in tabular formats (VCF, BCF, BGEN, etc.) and then compressed using off-the-shelf compression. In contrast, a GRG contains Mutation nodes (representing variants) and Sample nodes (representing haploid samples), where there is a path from a Mutation node to a Sample node if-and-only-if that sample contains that mutation. These paths go through internal nodes that represent common ancestry between multiple samples, and this can result in significant compression (30-50x smaller than .vcf.gz). Calculations on the whole dataset can be performed very quickly on GRG, using GRGL.

Recent releases (after v2.3) support the following improvements over the initial paper:

1. Graphs are more than 2x smaller (in RAM and on disk)
2. Graph construction is 10-25x faster
3. Loading graphs from disk is 10-20x faster
4. First-class matrix multiplication API matmul
5. (Prototype) unphased data is supported
6. GWAS, GWAS with covariates, PCA, and other analyses are available with grapp (pip install grapp)
7. Phenotype simulation is available with grg_pheno_sim (pip install grg_pheno_sim)
8. Construction from .vcf.gz now supports tabix indexes, making that input format feasible for large datasets
9. Better support for missing data, see the documentation

If you need to cite something, use "Enabling efficient analysis of biobank-scale data with genotype representation graphs".

Documentation

Check out the main documentation for core API documentation, examples, tutorials, etc. Things covered in the documentation include:

• Creating and using GRGs
• Performing GWAS, PCA, GWAS with covariates, or other analyses with GRG
• Simulating phenotypes with GRG
• Using GRG with Python (integration with numpy, pandas, scipy, etc.)

You can also download the tutorials as Jupyter Notebooks and work through them interactively.

Genotype Representation Graph Library (GRGL)

GRGL can be used as a library in both C++ and Python. Support is currently limited to Linux and MacOS. It contains both an API (see docs) and a set of command-line tools.

Installing from pip

If you just want to use the tools (e.g., constructing GRG or converting tree-sequence to GRG) and the Python API then you can install via pip (from PyPi).

pip install pygrgl

This will use prebuilt packages for most modern Linux situations, and will build from source for MacOS. In order to build from source it will require CMake (at least v3.14), zlib development headers, and a clang or GCC compiler that supports C++11.

Installing from conda

You can also install the conda package via the bioconda channel: conda install pygrgl.

Building (Python)

The Python installation installs the command line tools and Python libraries (the C++ executables are packaged as part of this). Make sure you clone with git clone --recursive!

Requires Python 3.7 or newer to be installed (including development headers). It is recommended that you build/install in a virtual environment.

python3 -m venv /path/to/MyEnv
source /path/to/MyEnv/bin/activate
python setup.py bdist_wheel               # Compiles C++, builds a wheel in the dist/ directory
pip install --force-reinstall dist/*.whl  # Install from wheel

Build and installation should take at most a few minutes on the typical computer. For more details on build options, see DEVELOPING.md.

Building (C++ only)

The C++ build is only necessary for folks who want to include GRGL as a library in their C++ project. Typically, you would include our CMake into your project via add_subdirectory, but you can also build standalone as below. Make sure you clone with git clone --recursive!

If you only intend to use GRGL from C++, you can just build it via CMake:

mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make -j4

See below to install the libraries to your system. It is recommended to install it to a custom location (prefix) since removing packages installed via make install is a pain otherwise. Example:

mkdir /path/to/grgl_installation/
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/path/to/grgl_installation/
make -j4
make install
# There should now be bin/, lib/, etc., directories under /path/to/grgl_installation/

Building (Docker)

We've included a Dockerfile if you want to use GRGL in a container.

Example to build:

docker build . -t grgl:latest

Example to run, constructing a GRG from an example VCF file:

docker run -v $PWD:/working -it grgl:latest bash -c "cd /working && grg construct --force /working/test/inputs/msprime.example.vcf"

Usage (Command line)

There is a command line tool that is mostly for file format conversion and performing common computations on the GRG. For more flexibility, use the Python or C++ APIs. After building and installing the Python version, run grg --help to see all the command options. Some examples are below.

Convert a tskit tree-sequence into a GRG. This creates my_arg_data.grg from my_arg_data.trees:

grg convert /path/to/my_arg_data.trees my_arg_data.grg

Load a GRG and emit some simple statistics about the GRG itself:

grg process stats my_arg_data.grg

To construct a GRG from a VCF file, use the grg construct command. (NOTE raw VCF is incredibly slow for non-trivial datasets, use BGZF indexed with tabix or IGD):

grg construct -j 1 path/to/foo.vcf.gz

To convert a VCF(.gz) to an IGD and then build a GRG:

pip install igdtools
igdtools path/to/foo.vcf -o foo.igd
grg construct -j 1 foo.igd

Increase -j to the number of threads you have. Construction for small datasets (such as those included as tests in this repository) should be very fast, on the order of seconds. Really large datasets (such as Biobank-scale whole genome sequences) can take on the order of hours when using lots of threads (e.g., 70). 1,000 Genomes Project chromosomes usually take on the order of a few minutes.

Usage (Python API)

See the provided jupyter notebooks and GettingStarted.md for more examples.

Limits

Quantity	Limit
Haploid samples	2,147,483,646
Total nodes	2,147,483,646
Total mutations (variants)	4,294,967,294
Total edges	18,446,744,073,709,551,615
Edges to/from a single node	4,294,967,295

Note: Node limits can theoretically be expanded to about a trillion, by turning on the LARGE_NODE_IDS preprocessor flag, but this mode is not well tested.