Dedicated hardware accelerators for the epistatic analysis of human genetic data

The recent advances in genomic microarrays design provide the possibility to retrieve hundreds of thousands of significative genetic features from patients at affordable costs. Understanding if non-linear interactions (epistatic relationships) between these features determine or not the arising of complex common multifactorial genetic diseases is a critical task for human geneticists. The algorithms able to detect such relationships, like the Multifactor Dimensionality Reduction (MDR) algorithm, are computationally expensive and their practical utility is very often limited by the amount of time required by the analysis. This paper presents three hardware-accelerated implementations of the MDR algorithm, tailored for many-core processors, Xilinx Virtex-5 FPGAs and generic GPUs respectively. These implementations provide timing performance improvements of up to two magnitude orders with respect to the software implementation.