Scalable analysis of large multi-ancestry biobanks by leveraging sparse ancestry-adjusted sample-relatedness

Linear mixed-effects models (LMMs) and ridge regression are commonly applied in genetic association studies to control for population structure and sample-relatedness. To control for sample-relatedness, the existing methods use empirical genetic relatedness matrices (GRM) either explicitly or conceptually. This works well with mostly homogeneous populations, however, in multi-ancestry heterogeneous populations, GRMs are confounded with population structure which leads to inflated type I error rates, massively increased computation, and reduced power. Here, we propose FastSparseGRM, a scalable pipeline for multi-ancestry Genome-Wide Association studies (GWAS) and Whole Genome Sequencing (WGS) studies. It utilizes a block-diagonal sparse ancestry-adjusted (BDSA) GRM to model sample-relatedness, and ancestry PCs as fixed effects to control for population structure. It is ~ 2540/4100/54 times faster than BOLT-LMM/fast-GWA/REGENIE for fitting the null LMM on 50,000 heterogeneous subjects. Through numerical simulations and both single-variant GWAS and rare variant WGS analyses of five biomarkers (Triglycerides, HDL, LDL, BMI, Total Bilirubin) on the entire UK Biobank data, we demonstrate that our approach scales to nearly half-a-million subjects and provides accurate p-value calibration and improved power compared to the existing methods.