TorchGWAS 1.0: GPU-accelerated GWAS at scale

Abstract Motivation Imaging, molecular and machine-learning workflows can generate thousands of quantitative phenotypes in one cohort. Repeating genome-wide scans trait by trait creates a computational and output bottleneck. Results TorchGWAS is a GPU-accelerated framework that uses batched operations for high-throughput, covariate-adjusted linear association testing across large panels of quantitative phenotypes. Across 500,036 allele-harmonized tests, its test statistics agreed with PLINK 2.0, with a Pearson correlation difference from one of 1.3 × 10⁻¹². On an NVIDIA H100 with 80 gigabytes of memory and a 48-core Intel Xeon Gold 6442Y host, median end-to-end runtimes for 4.57 billion associations spanning 8,931,083 variants, 512 phenotypes and 35,365 samples ranged from 28.46 to 58.95 seconds across four genotype formats. These measurements included writing 36.7 gigabytes of binary summary statistics. TorchGWAS supports efficient fixed-effect association screening at scale. Availability and implementation TorchGWAS is implemented in Python and distributed as a documented source repository at https://github.com/ZhiGroup/TorchGWAS.