MOWAS Multi-Omics Wide Association Study

MOWAS can exhaustively test potential epistasis between trillions of SNP pairs based on linear regression,Y~β1SNP1+β2SNP2+β3SNP1SNP2+covariates, where Y represents case-control status but is treated as a continuous variable to reduce logistic regression computational burden, and SNPs were coded as minor allele number plus one.  We test whether ß1 is significantly different from zero.  MOWAS was run in a heterogeneous supercomputing environment.  The analysis was programmed in Heterogeneous Interface for Portability (HIP) on the GPU and used the Radeon Open Computing Platform (ROCm) to invoke Basic Linear Algebra Subprograms (BLAS) to accelerate the matrix solving process.  Computational efficiency was optimized by using Multipoint Interface (MPI) master/slave mode to adaptively partition the computational tasks and allocate them to CPUs and GPUs on multiple compute nodes for asynchronous execution.  This analysis was performed using 80 compute nodes, 2560×86 CPU cores, and 320 GPU accelerators, and was completed in 21.3 hours with a double precision computational performance of 15.5 tera floating point operations per second (TFLOPS).