GWAS-Flow: A GPU accelerated framework for efficient permutation based genome-wide association studies

Motivation: Genome-wide association studies (GWAS) are one of the most commonly used methods to detect associations between complex traits and genomic polymorphisms. As both genotyping and phenotyping of large populations has become easier, typical modern GWAS have to cope with massive amounts of data. Thus, the computational demand for these analyses grew remarkably during the last decades. This is especially true, if one wants to implement permutation-based significance thresholds, instead of using the naive Bonferroni threshold. Permutation-based methods have the advantage to provide an adjusted multiple hypothesis correction threshold that takes the underlying phenotypic distribution into account and will thus remove the need to find the correct transformation for non Gaussian phenotypes. To enable efficient analyses of large datasets and the possibility to compute permutation-based significance thresholds, we used the machine learning framework TensorFlow to develop a linear mixed model (GWAS-Flow) that can make use of the available CPU or GPU infrastructure to decrease the time of the analyses especially for large datasets. Results: We were able to show that our application GWAS-Flow outperforms custom GWAS scripts in terms of speed without loosing accuracy. Apart from p-values, GWAS-Flow also computes summary statistics, such as the effect size and its standard error for each individual marker. The CPU-based version is the default choice for small data, while the GPU-based version of GWAS-Flow is especially suited for the analyses of big data. Availability: GWAS-Flow is freely available on GitHub (https://github.com/Joyvalley/GWAS_Flow) and is released under the terms of the MIT-License.