TorchLIMIX: GPU-accelerated multivariate genome-wide association studies

Abstract Summary We introduce TorchLIMIX, a GPU-accelerated PyTorch implementation of the LIMIX multivariate genome-wide association study pipeline. By leveraging batched GPU linear algebra, TorchLIMIX achieves speedups of up to two orders of magnitude over the original CPU-based implementation while maintaining numerically equivalent results and full concordance of significantly associated loci. In simulation studies, replacing the default initialization of the genetic covariance factor with a QR-based strategy reduces genomic inflation factors to near-unity values under the common and interaction effect null hypotheses, ensuring well-calibrated type I error control. Applying TorchLIMIX to metabolic traits of Arabidopsis thaliana measured in two experiments uncovered 37 additional associated SNPs at the same significance threshold used in the original univariate GWAS. Availability The TorchLIMIX pipeline is openly available on GitHub at https://github.com/bi-horn/torchLIMIX. An adapted version of the multivariate association testing part of the original LIMIX pipeline is available at https://github.com/bi-horn/LIMIX_modified. Contact bibiana.horn{at}hpi.de Supplementary information Supplementary materials are included with this submission. Competing Interest Statement The authors have declared no competing interest.