BioGPU: A High Performance Computing Tool for Genome-Wide Association Studies

The performance of computing tools for genome-wide association studies (GWAS) are measured by their computing speed, memory requirements, and statistical power. These three factors are determined by the statistical methods a tool implemented and how these methods are engineered to make full use of computer hardware resources. We developed a computing tool named BioGPU that implements a new statistical method, which is also referred to as BioGPU (Batched and Iterative Optimization for Genomic Probability Unification). BioGPU effectively controls false positives caused by population structure and unequal relatedness among individuals and improves statistical power - even when compared to mixed linear model methods. The BioGPU method requires much less complex computing time, nearly matching the linear computing time complexity of the general linear model. BioGPU was written in C computer language to maximize the capability of direct electronic circuit operations, including binary formatting of genotype input files and bit operations for matrix manipulations. To further increase computing speed, BioGPU was developed with parallel computational capacity, so that computing times decrease linearly with the number of central processing units. Furthermore, the parallel components are dissected small enough so that graphic processing units are also able to perform parallel computations. To solve the memory footprint bottleneck, BioGPU allows users to directly control memory usage when big data are analyzed on computers with limited memory. That is, users have the option to trade computing time for less memory usage. Based on these features above, BioGPU makes analyses of large and complex datasets feasible without supercomputers.