A Flexible Computational Framework Using R and Map-Reduce for Permutation Tests of Massive Genetic Analysis of...

In quantitative trait locus (QTL) mapping significance of putative QTL is often determined using permutation testing. The computational needs to calculate the significance level are immense, 10⁴ up to 10⁸ or even more permutations can be needed. We have previously introduced the PruneDIRECT algorithm for multiple QTL scan with epistatic interactions. This algorithm has specific strengths for permutation testing. Here, we present a flexible, parallel computing framework for identifying multiple interacting QTL using the PruneDIRECT algorithm which uses the map-reduce model as implemented in Hadoop. The framework is implemented in R, a widely used software tool among geneticists. This enables users to rearrange algorithmic steps to adapt genetic models, search algorithms, and parallelization steps to their needs in a flexible way. Our work underlines the maturity of accessing distributed parallel computing for computationally demanding bioinformatics applications through building workflows within existing scientific environments. We investigate the PruneDIRECT algorithm, comparing its performance to exhaustive search and DIRECT algorithm using our framework on a public cloud resource. We find that PruneDIRECT is vastly superior for permutation testing, and perform 2 ×10⁵ permutations for a 2D QTL problem in 15 hours, using 100 cloud processes. We show that our framework scales out almost linearly for a 3D QTL search.