Link to our final group presentation!
TODO: Fill in background on GWAS
G2WAS aims to develop a fast GWAS data analysis pipeline incorporating GPU acceleration and Machine learning.
rapids.ai will be the core SDK that powers this high performance compute. RAPIDS provides GPU-accelerated implementations of dataframe and machine learning algorithms which are tens to hundreds of times faster than CPU implementations.
Our high level goals for the hackathon include -
- Recreating the Hail GWAS sample using RAPIDS APIs
- Wrapping RAPIDS GWAS functions into high level APIs within gpu-gwas framework
- (if time permits) Scaling to larger VCF dataset (e.g. chr22)
Working doc for our team is here
To get the best experience running this framework (or developing it!), we recommend having access to a machine with the following build -
- Ubuntu 18.04+
- NVIDIA GPU (Pascal architecture onwards)
- NVIDIA CUDA Toolkit 10.1+ Instructions here
The primary dependency of this project is installing the RAPIDS frameworks. RAPIDS has very clear instructions for how to download and install it, based on the CUDA Toolkit, python version and operating system being used.
Our experiments were based on the following setup
# Make sure CUDA Toolkit 10.2 is installed
conda create -n rapids-0.16 -c rapidsai -c nvidia -c conda-forge \
-c defaults rapids=0.16 python=3.7 cudatoolkit=10.2
Once the conda environment is setup, additional python dependencies for gpugwas are installed from pip.
conda activate rapids-0.16
python3 -m pip install -r requirements.txt
To make sure your environment was setup properly, please run the following command from the repo root.
python workflow.py
The gpugwas package is broken up into multiple independent modules that deal with different components
of the GWAS pipeline. The modules are all located under the gpugwas folder.
gpugwas.io- This module contains I/O related functions such as loading a VCF/annotation file into a CUDA dataframe.gpugwas.filter- This module containts functions to filter out variants and samples and perform QC on the input data.gpugwas.algorithms- This module contains ML algorithm implementations in CUDA typically used in GWAS (e.g. linear regression, logistic regression, etc).gpugwas.viz- This module contains functions used in visualizing the GWAS model outputs (manhattan plots, q-q plots, etc)
Using the package components described above we have built a sample workflow that runs a toy GWAS example. The contents of this script follow the workflow outlined above.
We use the VCF and annotations dataset from the Hail GWAS example for our experiment. Please download them to your system first before proceeding with the sample.
Our workflow showcasing the relevant APIs and processing steps is detailed in jupyter notebook.
To run it on your system, please execute
jupyter notebook workflow.ipynb